2 * Generic hugetlb support.
3 * (C) Nadia Yvette Chambers, April 2004
5 #include <linux/list.h>
6 #include <linux/init.h>
8 #include <linux/seq_file.h>
9 #include <linux/sysctl.h>
10 #include <linux/highmem.h>
11 #include <linux/mmu_notifier.h>
12 #include <linux/nodemask.h>
13 #include <linux/pagemap.h>
14 #include <linux/mempolicy.h>
15 #include <linux/compiler.h>
16 #include <linux/cpuset.h>
17 #include <linux/mutex.h>
18 #include <linux/bootmem.h>
19 #include <linux/sysfs.h>
20 #include <linux/slab.h>
21 #include <linux/rmap.h>
22 #include <linux/swap.h>
23 #include <linux/swapops.h>
24 #include <linux/page-isolation.h>
25 #include <linux/jhash.h>
28 #include <asm/pgtable.h>
32 #include <linux/hugetlb.h>
33 #include <linux/hugetlb_cgroup.h>
34 #include <linux/node.h>
35 #include <linux/userfaultfd_k.h>
38 int hugepages_treat_as_movable
;
40 int hugetlb_max_hstate __read_mostly
;
41 unsigned int default_hstate_idx
;
42 struct hstate hstates
[HUGE_MAX_HSTATE
];
44 * Minimum page order among possible hugepage sizes, set to a proper value
47 static unsigned int minimum_order __read_mostly
= UINT_MAX
;
49 __initdata
LIST_HEAD(huge_boot_pages
);
51 /* for command line parsing */
52 static struct hstate
* __initdata parsed_hstate
;
53 static unsigned long __initdata default_hstate_max_huge_pages
;
54 static unsigned long __initdata default_hstate_size
;
55 static bool __initdata parsed_valid_hugepagesz
= true;
58 * Protects updates to hugepage_freelists, hugepage_activelist, nr_huge_pages,
59 * free_huge_pages, and surplus_huge_pages.
61 DEFINE_SPINLOCK(hugetlb_lock
);
64 * Serializes faults on the same logical page. This is used to
65 * prevent spurious OOMs when the hugepage pool is fully utilized.
67 static int num_fault_mutexes
;
68 struct mutex
*hugetlb_fault_mutex_table ____cacheline_aligned_in_smp
;
70 /* Forward declaration */
71 static int hugetlb_acct_memory(struct hstate
*h
, long delta
);
73 static inline void unlock_or_release_subpool(struct hugepage_subpool
*spool
)
75 bool free
= (spool
->count
== 0) && (spool
->used_hpages
== 0);
77 spin_unlock(&spool
->lock
);
79 /* If no pages are used, and no other handles to the subpool
80 * remain, give up any reservations mased on minimum size and
83 if (spool
->min_hpages
!= -1)
84 hugetlb_acct_memory(spool
->hstate
,
90 struct hugepage_subpool
*hugepage_new_subpool(struct hstate
*h
, long max_hpages
,
93 struct hugepage_subpool
*spool
;
95 spool
= kzalloc(sizeof(*spool
), GFP_KERNEL
);
99 spin_lock_init(&spool
->lock
);
101 spool
->max_hpages
= max_hpages
;
103 spool
->min_hpages
= min_hpages
;
105 if (min_hpages
!= -1 && hugetlb_acct_memory(h
, min_hpages
)) {
109 spool
->rsv_hpages
= min_hpages
;
114 void hugepage_put_subpool(struct hugepage_subpool
*spool
)
116 spin_lock(&spool
->lock
);
117 BUG_ON(!spool
->count
);
119 unlock_or_release_subpool(spool
);
123 * Subpool accounting for allocating and reserving pages.
124 * Return -ENOMEM if there are not enough resources to satisfy the
125 * the request. Otherwise, return the number of pages by which the
126 * global pools must be adjusted (upward). The returned value may
127 * only be different than the passed value (delta) in the case where
128 * a subpool minimum size must be manitained.
130 static long hugepage_subpool_get_pages(struct hugepage_subpool
*spool
,
138 spin_lock(&spool
->lock
);
140 if (spool
->max_hpages
!= -1) { /* maximum size accounting */
141 if ((spool
->used_hpages
+ delta
) <= spool
->max_hpages
)
142 spool
->used_hpages
+= delta
;
149 /* minimum size accounting */
150 if (spool
->min_hpages
!= -1 && spool
->rsv_hpages
) {
151 if (delta
> spool
->rsv_hpages
) {
153 * Asking for more reserves than those already taken on
154 * behalf of subpool. Return difference.
156 ret
= delta
- spool
->rsv_hpages
;
157 spool
->rsv_hpages
= 0;
159 ret
= 0; /* reserves already accounted for */
160 spool
->rsv_hpages
-= delta
;
165 spin_unlock(&spool
->lock
);
170 * Subpool accounting for freeing and unreserving pages.
171 * Return the number of global page reservations that must be dropped.
172 * The return value may only be different than the passed value (delta)
173 * in the case where a subpool minimum size must be maintained.
175 static long hugepage_subpool_put_pages(struct hugepage_subpool
*spool
,
183 spin_lock(&spool
->lock
);
185 if (spool
->max_hpages
!= -1) /* maximum size accounting */
186 spool
->used_hpages
-= delta
;
188 /* minimum size accounting */
189 if (spool
->min_hpages
!= -1 && spool
->used_hpages
< spool
->min_hpages
) {
190 if (spool
->rsv_hpages
+ delta
<= spool
->min_hpages
)
193 ret
= spool
->rsv_hpages
+ delta
- spool
->min_hpages
;
195 spool
->rsv_hpages
+= delta
;
196 if (spool
->rsv_hpages
> spool
->min_hpages
)
197 spool
->rsv_hpages
= spool
->min_hpages
;
201 * If hugetlbfs_put_super couldn't free spool due to an outstanding
202 * quota reference, free it now.
204 unlock_or_release_subpool(spool
);
209 static inline struct hugepage_subpool
*subpool_inode(struct inode
*inode
)
211 return HUGETLBFS_SB(inode
->i_sb
)->spool
;
214 static inline struct hugepage_subpool
*subpool_vma(struct vm_area_struct
*vma
)
216 return subpool_inode(file_inode(vma
->vm_file
));
220 * Region tracking -- allows tracking of reservations and instantiated pages
221 * across the pages in a mapping.
223 * The region data structures are embedded into a resv_map and protected
224 * by a resv_map's lock. The set of regions within the resv_map represent
225 * reservations for huge pages, or huge pages that have already been
226 * instantiated within the map. The from and to elements are huge page
227 * indicies into the associated mapping. from indicates the starting index
228 * of the region. to represents the first index past the end of the region.
230 * For example, a file region structure with from == 0 and to == 4 represents
231 * four huge pages in a mapping. It is important to note that the to element
232 * represents the first element past the end of the region. This is used in
233 * arithmetic as 4(to) - 0(from) = 4 huge pages in the region.
235 * Interval notation of the form [from, to) will be used to indicate that
236 * the endpoint from is inclusive and to is exclusive.
239 struct list_head link
;
245 * Add the huge page range represented by [f, t) to the reserve
246 * map. In the normal case, existing regions will be expanded
247 * to accommodate the specified range. Sufficient regions should
248 * exist for expansion due to the previous call to region_chg
249 * with the same range. However, it is possible that region_del
250 * could have been called after region_chg and modifed the map
251 * in such a way that no region exists to be expanded. In this
252 * case, pull a region descriptor from the cache associated with
253 * the map and use that for the new range.
255 * Return the number of new huge pages added to the map. This
256 * number is greater than or equal to zero.
258 static long region_add(struct resv_map
*resv
, long f
, long t
)
260 struct list_head
*head
= &resv
->regions
;
261 struct file_region
*rg
, *nrg
, *trg
;
264 spin_lock(&resv
->lock
);
265 /* Locate the region we are either in or before. */
266 list_for_each_entry(rg
, head
, link
)
271 * If no region exists which can be expanded to include the
272 * specified range, the list must have been modified by an
273 * interleving call to region_del(). Pull a region descriptor
274 * from the cache and use it for this range.
276 if (&rg
->link
== head
|| t
< rg
->from
) {
277 VM_BUG_ON(resv
->region_cache_count
<= 0);
279 resv
->region_cache_count
--;
280 nrg
= list_first_entry(&resv
->region_cache
, struct file_region
,
282 list_del(&nrg
->link
);
286 list_add(&nrg
->link
, rg
->link
.prev
);
292 /* Round our left edge to the current segment if it encloses us. */
296 /* Check for and consume any regions we now overlap with. */
298 list_for_each_entry_safe(rg
, trg
, rg
->link
.prev
, link
) {
299 if (&rg
->link
== head
)
304 /* If this area reaches higher then extend our area to
305 * include it completely. If this is not the first area
306 * which we intend to reuse, free it. */
310 /* Decrement return value by the deleted range.
311 * Another range will span this area so that by
312 * end of routine add will be >= zero
314 add
-= (rg
->to
- rg
->from
);
320 add
+= (nrg
->from
- f
); /* Added to beginning of region */
322 add
+= t
- nrg
->to
; /* Added to end of region */
326 resv
->adds_in_progress
--;
327 spin_unlock(&resv
->lock
);
333 * Examine the existing reserve map and determine how many
334 * huge pages in the specified range [f, t) are NOT currently
335 * represented. This routine is called before a subsequent
336 * call to region_add that will actually modify the reserve
337 * map to add the specified range [f, t). region_chg does
338 * not change the number of huge pages represented by the
339 * map. However, if the existing regions in the map can not
340 * be expanded to represent the new range, a new file_region
341 * structure is added to the map as a placeholder. This is
342 * so that the subsequent region_add call will have all the
343 * regions it needs and will not fail.
345 * Upon entry, region_chg will also examine the cache of region descriptors
346 * associated with the map. If there are not enough descriptors cached, one
347 * will be allocated for the in progress add operation.
349 * Returns the number of huge pages that need to be added to the existing
350 * reservation map for the range [f, t). This number is greater or equal to
351 * zero. -ENOMEM is returned if a new file_region structure or cache entry
352 * is needed and can not be allocated.
354 static long region_chg(struct resv_map
*resv
, long f
, long t
)
356 struct list_head
*head
= &resv
->regions
;
357 struct file_region
*rg
, *nrg
= NULL
;
361 spin_lock(&resv
->lock
);
363 resv
->adds_in_progress
++;
366 * Check for sufficient descriptors in the cache to accommodate
367 * the number of in progress add operations.
369 if (resv
->adds_in_progress
> resv
->region_cache_count
) {
370 struct file_region
*trg
;
372 VM_BUG_ON(resv
->adds_in_progress
- resv
->region_cache_count
> 1);
373 /* Must drop lock to allocate a new descriptor. */
374 resv
->adds_in_progress
--;
375 spin_unlock(&resv
->lock
);
377 trg
= kmalloc(sizeof(*trg
), GFP_KERNEL
);
383 spin_lock(&resv
->lock
);
384 list_add(&trg
->link
, &resv
->region_cache
);
385 resv
->region_cache_count
++;
389 /* Locate the region we are before or in. */
390 list_for_each_entry(rg
, head
, link
)
394 /* If we are below the current region then a new region is required.
395 * Subtle, allocate a new region at the position but make it zero
396 * size such that we can guarantee to record the reservation. */
397 if (&rg
->link
== head
|| t
< rg
->from
) {
399 resv
->adds_in_progress
--;
400 spin_unlock(&resv
->lock
);
401 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
407 INIT_LIST_HEAD(&nrg
->link
);
411 list_add(&nrg
->link
, rg
->link
.prev
);
416 /* Round our left edge to the current segment if it encloses us. */
421 /* Check for and consume any regions we now overlap with. */
422 list_for_each_entry(rg
, rg
->link
.prev
, link
) {
423 if (&rg
->link
== head
)
428 /* We overlap with this area, if it extends further than
429 * us then we must extend ourselves. Account for its
430 * existing reservation. */
435 chg
-= rg
->to
- rg
->from
;
439 spin_unlock(&resv
->lock
);
440 /* We already know we raced and no longer need the new region */
444 spin_unlock(&resv
->lock
);
449 * Abort the in progress add operation. The adds_in_progress field
450 * of the resv_map keeps track of the operations in progress between
451 * calls to region_chg and region_add. Operations are sometimes
452 * aborted after the call to region_chg. In such cases, region_abort
453 * is called to decrement the adds_in_progress counter.
455 * NOTE: The range arguments [f, t) are not needed or used in this
456 * routine. They are kept to make reading the calling code easier as
457 * arguments will match the associated region_chg call.
459 static void region_abort(struct resv_map
*resv
, long f
, long t
)
461 spin_lock(&resv
->lock
);
462 VM_BUG_ON(!resv
->region_cache_count
);
463 resv
->adds_in_progress
--;
464 spin_unlock(&resv
->lock
);
468 * Delete the specified range [f, t) from the reserve map. If the
469 * t parameter is LONG_MAX, this indicates that ALL regions after f
470 * should be deleted. Locate the regions which intersect [f, t)
471 * and either trim, delete or split the existing regions.
473 * Returns the number of huge pages deleted from the reserve map.
474 * In the normal case, the return value is zero or more. In the
475 * case where a region must be split, a new region descriptor must
476 * be allocated. If the allocation fails, -ENOMEM will be returned.
477 * NOTE: If the parameter t == LONG_MAX, then we will never split
478 * a region and possibly return -ENOMEM. Callers specifying
479 * t == LONG_MAX do not need to check for -ENOMEM error.
481 static long region_del(struct resv_map
*resv
, long f
, long t
)
483 struct list_head
*head
= &resv
->regions
;
484 struct file_region
*rg
, *trg
;
485 struct file_region
*nrg
= NULL
;
489 spin_lock(&resv
->lock
);
490 list_for_each_entry_safe(rg
, trg
, head
, link
) {
492 * Skip regions before the range to be deleted. file_region
493 * ranges are normally of the form [from, to). However, there
494 * may be a "placeholder" entry in the map which is of the form
495 * (from, to) with from == to. Check for placeholder entries
496 * at the beginning of the range to be deleted.
498 if (rg
->to
<= f
&& (rg
->to
!= rg
->from
|| rg
->to
!= f
))
504 if (f
> rg
->from
&& t
< rg
->to
) { /* Must split region */
506 * Check for an entry in the cache before dropping
507 * lock and attempting allocation.
510 resv
->region_cache_count
> resv
->adds_in_progress
) {
511 nrg
= list_first_entry(&resv
->region_cache
,
514 list_del(&nrg
->link
);
515 resv
->region_cache_count
--;
519 spin_unlock(&resv
->lock
);
520 nrg
= kmalloc(sizeof(*nrg
), GFP_KERNEL
);
528 /* New entry for end of split region */
531 INIT_LIST_HEAD(&nrg
->link
);
533 /* Original entry is trimmed */
536 list_add(&nrg
->link
, &rg
->link
);
541 if (f
<= rg
->from
&& t
>= rg
->to
) { /* Remove entire region */
542 del
+= rg
->to
- rg
->from
;
548 if (f
<= rg
->from
) { /* Trim beginning of region */
551 } else { /* Trim end of region */
557 spin_unlock(&resv
->lock
);
563 * A rare out of memory error was encountered which prevented removal of
564 * the reserve map region for a page. The huge page itself was free'ed
565 * and removed from the page cache. This routine will adjust the subpool
566 * usage count, and the global reserve count if needed. By incrementing
567 * these counts, the reserve map entry which could not be deleted will
568 * appear as a "reserved" entry instead of simply dangling with incorrect
571 void hugetlb_fix_reserve_counts(struct inode
*inode
)
573 struct hugepage_subpool
*spool
= subpool_inode(inode
);
576 rsv_adjust
= hugepage_subpool_get_pages(spool
, 1);
578 struct hstate
*h
= hstate_inode(inode
);
580 hugetlb_acct_memory(h
, 1);
585 * Count and return the number of huge pages in the reserve map
586 * that intersect with the range [f, t).
588 static long region_count(struct resv_map
*resv
, long f
, long t
)
590 struct list_head
*head
= &resv
->regions
;
591 struct file_region
*rg
;
594 spin_lock(&resv
->lock
);
595 /* Locate each segment we overlap with, and count that overlap. */
596 list_for_each_entry(rg
, head
, link
) {
605 seg_from
= max(rg
->from
, f
);
606 seg_to
= min(rg
->to
, t
);
608 chg
+= seg_to
- seg_from
;
610 spin_unlock(&resv
->lock
);
616 * Convert the address within this vma to the page offset within
617 * the mapping, in pagecache page units; huge pages here.
619 static pgoff_t
vma_hugecache_offset(struct hstate
*h
,
620 struct vm_area_struct
*vma
, unsigned long address
)
622 return ((address
- vma
->vm_start
) >> huge_page_shift(h
)) +
623 (vma
->vm_pgoff
>> huge_page_order(h
));
626 pgoff_t
linear_hugepage_index(struct vm_area_struct
*vma
,
627 unsigned long address
)
629 return vma_hugecache_offset(hstate_vma(vma
), vma
, address
);
631 EXPORT_SYMBOL_GPL(linear_hugepage_index
);
634 * Return the size of the pages allocated when backing a VMA. In the majority
635 * cases this will be same size as used by the page table entries.
637 unsigned long vma_kernel_pagesize(struct vm_area_struct
*vma
)
639 struct hstate
*hstate
;
641 if (!is_vm_hugetlb_page(vma
))
644 hstate
= hstate_vma(vma
);
646 return 1UL << huge_page_shift(hstate
);
648 EXPORT_SYMBOL_GPL(vma_kernel_pagesize
);
651 * Return the page size being used by the MMU to back a VMA. In the majority
652 * of cases, the page size used by the kernel matches the MMU size. On
653 * architectures where it differs, an architecture-specific version of this
654 * function is required.
656 #ifndef vma_mmu_pagesize
657 unsigned long vma_mmu_pagesize(struct vm_area_struct
*vma
)
659 return vma_kernel_pagesize(vma
);
664 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
665 * bits of the reservation map pointer, which are always clear due to
668 #define HPAGE_RESV_OWNER (1UL << 0)
669 #define HPAGE_RESV_UNMAPPED (1UL << 1)
670 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
673 * These helpers are used to track how many pages are reserved for
674 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
675 * is guaranteed to have their future faults succeed.
677 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
678 * the reserve counters are updated with the hugetlb_lock held. It is safe
679 * to reset the VMA at fork() time as it is not in use yet and there is no
680 * chance of the global counters getting corrupted as a result of the values.
682 * The private mapping reservation is represented in a subtly different
683 * manner to a shared mapping. A shared mapping has a region map associated
684 * with the underlying file, this region map represents the backing file
685 * pages which have ever had a reservation assigned which this persists even
686 * after the page is instantiated. A private mapping has a region map
687 * associated with the original mmap which is attached to all VMAs which
688 * reference it, this region map represents those offsets which have consumed
689 * reservation ie. where pages have been instantiated.
691 static unsigned long get_vma_private_data(struct vm_area_struct
*vma
)
693 return (unsigned long)vma
->vm_private_data
;
696 static void set_vma_private_data(struct vm_area_struct
*vma
,
699 vma
->vm_private_data
= (void *)value
;
702 struct resv_map
*resv_map_alloc(void)
704 struct resv_map
*resv_map
= kmalloc(sizeof(*resv_map
), GFP_KERNEL
);
705 struct file_region
*rg
= kmalloc(sizeof(*rg
), GFP_KERNEL
);
707 if (!resv_map
|| !rg
) {
713 kref_init(&resv_map
->refs
);
714 spin_lock_init(&resv_map
->lock
);
715 INIT_LIST_HEAD(&resv_map
->regions
);
717 resv_map
->adds_in_progress
= 0;
719 INIT_LIST_HEAD(&resv_map
->region_cache
);
720 list_add(&rg
->link
, &resv_map
->region_cache
);
721 resv_map
->region_cache_count
= 1;
726 void resv_map_release(struct kref
*ref
)
728 struct resv_map
*resv_map
= container_of(ref
, struct resv_map
, refs
);
729 struct list_head
*head
= &resv_map
->region_cache
;
730 struct file_region
*rg
, *trg
;
732 /* Clear out any active regions before we release the map. */
733 region_del(resv_map
, 0, LONG_MAX
);
735 /* ... and any entries left in the cache */
736 list_for_each_entry_safe(rg
, trg
, head
, link
) {
741 VM_BUG_ON(resv_map
->adds_in_progress
);
746 static inline struct resv_map
*inode_resv_map(struct inode
*inode
)
748 return inode
->i_mapping
->private_data
;
751 static struct resv_map
*vma_resv_map(struct vm_area_struct
*vma
)
753 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
754 if (vma
->vm_flags
& VM_MAYSHARE
) {
755 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
756 struct inode
*inode
= mapping
->host
;
758 return inode_resv_map(inode
);
761 return (struct resv_map
*)(get_vma_private_data(vma
) &
766 static void set_vma_resv_map(struct vm_area_struct
*vma
, struct resv_map
*map
)
768 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
769 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
771 set_vma_private_data(vma
, (get_vma_private_data(vma
) &
772 HPAGE_RESV_MASK
) | (unsigned long)map
);
775 static void set_vma_resv_flags(struct vm_area_struct
*vma
, unsigned long flags
)
777 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
778 VM_BUG_ON_VMA(vma
->vm_flags
& VM_MAYSHARE
, vma
);
780 set_vma_private_data(vma
, get_vma_private_data(vma
) | flags
);
783 static int is_vma_resv_set(struct vm_area_struct
*vma
, unsigned long flag
)
785 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
787 return (get_vma_private_data(vma
) & flag
) != 0;
790 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
791 void reset_vma_resv_huge_pages(struct vm_area_struct
*vma
)
793 VM_BUG_ON_VMA(!is_vm_hugetlb_page(vma
), vma
);
794 if (!(vma
->vm_flags
& VM_MAYSHARE
))
795 vma
->vm_private_data
= (void *)0;
798 /* Returns true if the VMA has associated reserve pages */
799 static bool vma_has_reserves(struct vm_area_struct
*vma
, long chg
)
801 if (vma
->vm_flags
& VM_NORESERVE
) {
803 * This address is already reserved by other process(chg == 0),
804 * so, we should decrement reserved count. Without decrementing,
805 * reserve count remains after releasing inode, because this
806 * allocated page will go into page cache and is regarded as
807 * coming from reserved pool in releasing step. Currently, we
808 * don't have any other solution to deal with this situation
809 * properly, so add work-around here.
811 if (vma
->vm_flags
& VM_MAYSHARE
&& chg
== 0)
817 /* Shared mappings always use reserves */
818 if (vma
->vm_flags
& VM_MAYSHARE
) {
820 * We know VM_NORESERVE is not set. Therefore, there SHOULD
821 * be a region map for all pages. The only situation where
822 * there is no region map is if a hole was punched via
823 * fallocate. In this case, there really are no reverves to
824 * use. This situation is indicated if chg != 0.
833 * Only the process that called mmap() has reserves for
836 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
)) {
838 * Like the shared case above, a hole punch or truncate
839 * could have been performed on the private mapping.
840 * Examine the value of chg to determine if reserves
841 * actually exist or were previously consumed.
842 * Very Subtle - The value of chg comes from a previous
843 * call to vma_needs_reserves(). The reserve map for
844 * private mappings has different (opposite) semantics
845 * than that of shared mappings. vma_needs_reserves()
846 * has already taken this difference in semantics into
847 * account. Therefore, the meaning of chg is the same
848 * as in the shared case above. Code could easily be
849 * combined, but keeping it separate draws attention to
850 * subtle differences.
861 static void enqueue_huge_page(struct hstate
*h
, struct page
*page
)
863 int nid
= page_to_nid(page
);
864 list_move(&page
->lru
, &h
->hugepage_freelists
[nid
]);
865 h
->free_huge_pages
++;
866 h
->free_huge_pages_node
[nid
]++;
869 static struct page
*dequeue_huge_page_node(struct hstate
*h
, int nid
)
873 list_for_each_entry(page
, &h
->hugepage_freelists
[nid
], lru
)
874 if (!is_migrate_isolate_page(page
))
877 * if 'non-isolated free hugepage' not found on the list,
878 * the allocation fails.
880 if (&h
->hugepage_freelists
[nid
] == &page
->lru
)
882 list_move(&page
->lru
, &h
->hugepage_activelist
);
883 set_page_refcounted(page
);
884 h
->free_huge_pages
--;
885 h
->free_huge_pages_node
[nid
]--;
889 /* Movability of hugepages depends on migration support. */
890 static inline gfp_t
htlb_alloc_mask(struct hstate
*h
)
892 if (hugepages_treat_as_movable
|| hugepage_migration_supported(h
))
893 return GFP_HIGHUSER_MOVABLE
;
898 static struct page
*dequeue_huge_page_vma(struct hstate
*h
,
899 struct vm_area_struct
*vma
,
900 unsigned long address
, int avoid_reserve
,
903 struct page
*page
= NULL
;
904 struct mempolicy
*mpol
;
905 nodemask_t
*nodemask
;
906 struct zonelist
*zonelist
;
909 unsigned int cpuset_mems_cookie
;
912 * A child process with MAP_PRIVATE mappings created by their parent
913 * have no page reserves. This check ensures that reservations are
914 * not "stolen". The child may still get SIGKILLed
916 if (!vma_has_reserves(vma
, chg
) &&
917 h
->free_huge_pages
- h
->resv_huge_pages
== 0)
920 /* If reserves cannot be used, ensure enough pages are in the pool */
921 if (avoid_reserve
&& h
->free_huge_pages
- h
->resv_huge_pages
== 0)
925 cpuset_mems_cookie
= read_mems_allowed_begin();
926 zonelist
= huge_zonelist(vma
, address
,
927 htlb_alloc_mask(h
), &mpol
, &nodemask
);
929 for_each_zone_zonelist_nodemask(zone
, z
, zonelist
,
930 MAX_NR_ZONES
- 1, nodemask
) {
931 if (cpuset_zone_allowed(zone
, htlb_alloc_mask(h
))) {
932 page
= dequeue_huge_page_node(h
, zone_to_nid(zone
));
936 if (!vma_has_reserves(vma
, chg
))
939 SetPagePrivate(page
);
940 h
->resv_huge_pages
--;
947 if (unlikely(!page
&& read_mems_allowed_retry(cpuset_mems_cookie
)))
956 * common helper functions for hstate_next_node_to_{alloc|free}.
957 * We may have allocated or freed a huge page based on a different
958 * nodes_allowed previously, so h->next_node_to_{alloc|free} might
959 * be outside of *nodes_allowed. Ensure that we use an allowed
960 * node for alloc or free.
962 static int next_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
964 nid
= next_node_in(nid
, *nodes_allowed
);
965 VM_BUG_ON(nid
>= MAX_NUMNODES
);
970 static int get_valid_node_allowed(int nid
, nodemask_t
*nodes_allowed
)
972 if (!node_isset(nid
, *nodes_allowed
))
973 nid
= next_node_allowed(nid
, nodes_allowed
);
978 * returns the previously saved node ["this node"] from which to
979 * allocate a persistent huge page for the pool and advance the
980 * next node from which to allocate, handling wrap at end of node
983 static int hstate_next_node_to_alloc(struct hstate
*h
,
984 nodemask_t
*nodes_allowed
)
988 VM_BUG_ON(!nodes_allowed
);
990 nid
= get_valid_node_allowed(h
->next_nid_to_alloc
, nodes_allowed
);
991 h
->next_nid_to_alloc
= next_node_allowed(nid
, nodes_allowed
);
997 * helper for free_pool_huge_page() - return the previously saved
998 * node ["this node"] from which to free a huge page. Advance the
999 * next node id whether or not we find a free huge page to free so
1000 * that the next attempt to free addresses the next node.
1002 static int hstate_next_node_to_free(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1006 VM_BUG_ON(!nodes_allowed
);
1008 nid
= get_valid_node_allowed(h
->next_nid_to_free
, nodes_allowed
);
1009 h
->next_nid_to_free
= next_node_allowed(nid
, nodes_allowed
);
1014 #define for_each_node_mask_to_alloc(hs, nr_nodes, node, mask) \
1015 for (nr_nodes = nodes_weight(*mask); \
1017 ((node = hstate_next_node_to_alloc(hs, mask)) || 1); \
1020 #define for_each_node_mask_to_free(hs, nr_nodes, node, mask) \
1021 for (nr_nodes = nodes_weight(*mask); \
1023 ((node = hstate_next_node_to_free(hs, mask)) || 1); \
1026 #if defined(CONFIG_ARCH_HAS_GIGANTIC_PAGE) && \
1027 ((defined(CONFIG_MEMORY_ISOLATION) && defined(CONFIG_COMPACTION)) || \
1028 defined(CONFIG_CMA))
1029 static void destroy_compound_gigantic_page(struct page
*page
,
1033 int nr_pages
= 1 << order
;
1034 struct page
*p
= page
+ 1;
1036 atomic_set(compound_mapcount_ptr(page
), 0);
1037 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1038 clear_compound_head(p
);
1039 set_page_refcounted(p
);
1042 set_compound_order(page
, 0);
1043 __ClearPageHead(page
);
1046 static void free_gigantic_page(struct page
*page
, unsigned int order
)
1048 free_contig_range(page_to_pfn(page
), 1 << order
);
1051 static int __alloc_gigantic_page(unsigned long start_pfn
,
1052 unsigned long nr_pages
)
1054 unsigned long end_pfn
= start_pfn
+ nr_pages
;
1055 return alloc_contig_range(start_pfn
, end_pfn
, MIGRATE_MOVABLE
,
1059 static bool pfn_range_valid_gigantic(struct zone
*z
,
1060 unsigned long start_pfn
, unsigned long nr_pages
)
1062 unsigned long i
, end_pfn
= start_pfn
+ nr_pages
;
1065 for (i
= start_pfn
; i
< end_pfn
; i
++) {
1069 page
= pfn_to_page(i
);
1071 if (page_zone(page
) != z
)
1074 if (PageReserved(page
))
1077 if (page_count(page
) > 0)
1087 static bool zone_spans_last_pfn(const struct zone
*zone
,
1088 unsigned long start_pfn
, unsigned long nr_pages
)
1090 unsigned long last_pfn
= start_pfn
+ nr_pages
- 1;
1091 return zone_spans_pfn(zone
, last_pfn
);
1094 static struct page
*alloc_gigantic_page(int nid
, unsigned int order
)
1096 unsigned long nr_pages
= 1 << order
;
1097 unsigned long ret
, pfn
, flags
;
1100 z
= NODE_DATA(nid
)->node_zones
;
1101 for (; z
- NODE_DATA(nid
)->node_zones
< MAX_NR_ZONES
; z
++) {
1102 spin_lock_irqsave(&z
->lock
, flags
);
1104 pfn
= ALIGN(z
->zone_start_pfn
, nr_pages
);
1105 while (zone_spans_last_pfn(z
, pfn
, nr_pages
)) {
1106 if (pfn_range_valid_gigantic(z
, pfn
, nr_pages
)) {
1108 * We release the zone lock here because
1109 * alloc_contig_range() will also lock the zone
1110 * at some point. If there's an allocation
1111 * spinning on this lock, it may win the race
1112 * and cause alloc_contig_range() to fail...
1114 spin_unlock_irqrestore(&z
->lock
, flags
);
1115 ret
= __alloc_gigantic_page(pfn
, nr_pages
);
1117 return pfn_to_page(pfn
);
1118 spin_lock_irqsave(&z
->lock
, flags
);
1123 spin_unlock_irqrestore(&z
->lock
, flags
);
1129 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
);
1130 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
);
1132 static struct page
*alloc_fresh_gigantic_page_node(struct hstate
*h
, int nid
)
1136 page
= alloc_gigantic_page(nid
, huge_page_order(h
));
1138 prep_compound_gigantic_page(page
, huge_page_order(h
));
1139 prep_new_huge_page(h
, page
, nid
);
1145 static int alloc_fresh_gigantic_page(struct hstate
*h
,
1146 nodemask_t
*nodes_allowed
)
1148 struct page
*page
= NULL
;
1151 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1152 page
= alloc_fresh_gigantic_page_node(h
, node
);
1160 static inline bool gigantic_page_supported(void) { return true; }
1162 static inline bool gigantic_page_supported(void) { return false; }
1163 static inline void free_gigantic_page(struct page
*page
, unsigned int order
) { }
1164 static inline void destroy_compound_gigantic_page(struct page
*page
,
1165 unsigned int order
) { }
1166 static inline int alloc_fresh_gigantic_page(struct hstate
*h
,
1167 nodemask_t
*nodes_allowed
) { return 0; }
1170 static void update_and_free_page(struct hstate
*h
, struct page
*page
)
1174 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
1178 h
->nr_huge_pages_node
[page_to_nid(page
)]--;
1179 for (i
= 0; i
< pages_per_huge_page(h
); i
++) {
1180 page
[i
].flags
&= ~(1 << PG_locked
| 1 << PG_error
|
1181 1 << PG_referenced
| 1 << PG_dirty
|
1182 1 << PG_active
| 1 << PG_private
|
1185 VM_BUG_ON_PAGE(hugetlb_cgroup_from_page(page
), page
);
1186 set_compound_page_dtor(page
, NULL_COMPOUND_DTOR
);
1187 set_page_refcounted(page
);
1188 if (hstate_is_gigantic(h
)) {
1189 destroy_compound_gigantic_page(page
, huge_page_order(h
));
1190 free_gigantic_page(page
, huge_page_order(h
));
1192 __free_pages(page
, huge_page_order(h
));
1196 struct hstate
*size_to_hstate(unsigned long size
)
1200 for_each_hstate(h
) {
1201 if (huge_page_size(h
) == size
)
1208 * Test to determine whether the hugepage is "active/in-use" (i.e. being linked
1209 * to hstate->hugepage_activelist.)
1211 * This function can be called for tail pages, but never returns true for them.
1213 bool page_huge_active(struct page
*page
)
1215 VM_BUG_ON_PAGE(!PageHuge(page
), page
);
1216 return PageHead(page
) && PagePrivate(&page
[1]);
1219 /* never called for tail page */
1220 static void set_page_huge_active(struct page
*page
)
1222 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1223 SetPagePrivate(&page
[1]);
1226 static void clear_page_huge_active(struct page
*page
)
1228 VM_BUG_ON_PAGE(!PageHeadHuge(page
), page
);
1229 ClearPagePrivate(&page
[1]);
1232 void free_huge_page(struct page
*page
)
1235 * Can't pass hstate in here because it is called from the
1236 * compound page destructor.
1238 struct hstate
*h
= page_hstate(page
);
1239 int nid
= page_to_nid(page
);
1240 struct hugepage_subpool
*spool
=
1241 (struct hugepage_subpool
*)page_private(page
);
1242 bool restore_reserve
;
1244 set_page_private(page
, 0);
1245 page
->mapping
= NULL
;
1246 VM_BUG_ON_PAGE(page_count(page
), page
);
1247 VM_BUG_ON_PAGE(page_mapcount(page
), page
);
1248 restore_reserve
= PagePrivate(page
);
1249 ClearPagePrivate(page
);
1252 * A return code of zero implies that the subpool will be under its
1253 * minimum size if the reservation is not restored after page is free.
1254 * Therefore, force restore_reserve operation.
1256 if (hugepage_subpool_put_pages(spool
, 1) == 0)
1257 restore_reserve
= true;
1259 spin_lock(&hugetlb_lock
);
1260 clear_page_huge_active(page
);
1261 hugetlb_cgroup_uncharge_page(hstate_index(h
),
1262 pages_per_huge_page(h
), page
);
1263 if (restore_reserve
)
1264 h
->resv_huge_pages
++;
1266 if (h
->surplus_huge_pages_node
[nid
]) {
1267 /* remove the page from active list */
1268 list_del(&page
->lru
);
1269 update_and_free_page(h
, page
);
1270 h
->surplus_huge_pages
--;
1271 h
->surplus_huge_pages_node
[nid
]--;
1273 arch_clear_hugepage_flags(page
);
1274 enqueue_huge_page(h
, page
);
1276 spin_unlock(&hugetlb_lock
);
1279 static void prep_new_huge_page(struct hstate
*h
, struct page
*page
, int nid
)
1281 INIT_LIST_HEAD(&page
->lru
);
1282 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1283 spin_lock(&hugetlb_lock
);
1284 set_hugetlb_cgroup(page
, NULL
);
1286 h
->nr_huge_pages_node
[nid
]++;
1287 spin_unlock(&hugetlb_lock
);
1288 put_page(page
); /* free it into the hugepage allocator */
1291 static void prep_compound_gigantic_page(struct page
*page
, unsigned int order
)
1294 int nr_pages
= 1 << order
;
1295 struct page
*p
= page
+ 1;
1297 /* we rely on prep_new_huge_page to set the destructor */
1298 set_compound_order(page
, order
);
1299 __ClearPageReserved(page
);
1300 __SetPageHead(page
);
1301 for (i
= 1; i
< nr_pages
; i
++, p
= mem_map_next(p
, page
, i
)) {
1303 * For gigantic hugepages allocated through bootmem at
1304 * boot, it's safer to be consistent with the not-gigantic
1305 * hugepages and clear the PG_reserved bit from all tail pages
1306 * too. Otherwse drivers using get_user_pages() to access tail
1307 * pages may get the reference counting wrong if they see
1308 * PG_reserved set on a tail page (despite the head page not
1309 * having PG_reserved set). Enforcing this consistency between
1310 * head and tail pages allows drivers to optimize away a check
1311 * on the head page when they need know if put_page() is needed
1312 * after get_user_pages().
1314 __ClearPageReserved(p
);
1315 set_page_count(p
, 0);
1316 set_compound_head(p
, page
);
1318 atomic_set(compound_mapcount_ptr(page
), -1);
1322 * PageHuge() only returns true for hugetlbfs pages, but not for normal or
1323 * transparent huge pages. See the PageTransHuge() documentation for more
1326 int PageHuge(struct page
*page
)
1328 if (!PageCompound(page
))
1331 page
= compound_head(page
);
1332 return page
[1].compound_dtor
== HUGETLB_PAGE_DTOR
;
1334 EXPORT_SYMBOL_GPL(PageHuge
);
1337 * PageHeadHuge() only returns true for hugetlbfs head page, but not for
1338 * normal or transparent huge pages.
1340 int PageHeadHuge(struct page
*page_head
)
1342 if (!PageHead(page_head
))
1345 return get_compound_page_dtor(page_head
) == free_huge_page
;
1348 pgoff_t
__basepage_index(struct page
*page
)
1350 struct page
*page_head
= compound_head(page
);
1351 pgoff_t index
= page_index(page_head
);
1352 unsigned long compound_idx
;
1354 if (!PageHuge(page_head
))
1355 return page_index(page
);
1357 if (compound_order(page_head
) >= MAX_ORDER
)
1358 compound_idx
= page_to_pfn(page
) - page_to_pfn(page_head
);
1360 compound_idx
= page
- page_head
;
1362 return (index
<< compound_order(page_head
)) + compound_idx
;
1365 static struct page
*alloc_fresh_huge_page_node(struct hstate
*h
, int nid
)
1369 page
= __alloc_pages_node(nid
,
1370 htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_THISNODE
|
1371 __GFP_REPEAT
|__GFP_NOWARN
,
1372 huge_page_order(h
));
1374 prep_new_huge_page(h
, page
, nid
);
1380 static int alloc_fresh_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
)
1386 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
1387 page
= alloc_fresh_huge_page_node(h
, node
);
1395 count_vm_event(HTLB_BUDDY_PGALLOC
);
1397 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1403 * Free huge page from pool from next node to free.
1404 * Attempt to keep persistent huge pages more or less
1405 * balanced over allowed nodes.
1406 * Called with hugetlb_lock locked.
1408 static int free_pool_huge_page(struct hstate
*h
, nodemask_t
*nodes_allowed
,
1414 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
1416 * If we're returning unused surplus pages, only examine
1417 * nodes with surplus pages.
1419 if ((!acct_surplus
|| h
->surplus_huge_pages_node
[node
]) &&
1420 !list_empty(&h
->hugepage_freelists
[node
])) {
1422 list_entry(h
->hugepage_freelists
[node
].next
,
1424 list_del(&page
->lru
);
1425 h
->free_huge_pages
--;
1426 h
->free_huge_pages_node
[node
]--;
1428 h
->surplus_huge_pages
--;
1429 h
->surplus_huge_pages_node
[node
]--;
1431 update_and_free_page(h
, page
);
1441 * Dissolve a given free hugepage into free buddy pages. This function does
1442 * nothing for in-use (including surplus) hugepages. Returns -EBUSY if the
1443 * number of free hugepages would be reduced below the number of reserved
1446 static int dissolve_free_huge_page(struct page
*page
)
1450 spin_lock(&hugetlb_lock
);
1451 if (PageHuge(page
) && !page_count(page
)) {
1452 struct page
*head
= compound_head(page
);
1453 struct hstate
*h
= page_hstate(head
);
1454 int nid
= page_to_nid(head
);
1455 if (h
->free_huge_pages
- h
->resv_huge_pages
== 0) {
1459 list_del(&head
->lru
);
1460 h
->free_huge_pages
--;
1461 h
->free_huge_pages_node
[nid
]--;
1462 h
->max_huge_pages
--;
1463 update_and_free_page(h
, head
);
1466 spin_unlock(&hugetlb_lock
);
1471 * Dissolve free hugepages in a given pfn range. Used by memory hotplug to
1472 * make specified memory blocks removable from the system.
1473 * Note that this will dissolve a free gigantic hugepage completely, if any
1474 * part of it lies within the given range.
1475 * Also note that if dissolve_free_huge_page() returns with an error, all
1476 * free hugepages that were dissolved before that error are lost.
1478 int dissolve_free_huge_pages(unsigned long start_pfn
, unsigned long end_pfn
)
1484 if (!hugepages_supported())
1487 for (pfn
= start_pfn
; pfn
< end_pfn
; pfn
+= 1 << minimum_order
) {
1488 page
= pfn_to_page(pfn
);
1489 if (PageHuge(page
) && !page_count(page
)) {
1490 rc
= dissolve_free_huge_page(page
);
1500 * There are 3 ways this can get called:
1501 * 1. With vma+addr: we use the VMA's memory policy
1502 * 2. With !vma, but nid=NUMA_NO_NODE: We try to allocate a huge
1503 * page from any node, and let the buddy allocator itself figure
1505 * 3. With !vma, but nid!=NUMA_NO_NODE. We allocate a huge page
1506 * strictly from 'nid'
1508 static struct page
*__hugetlb_alloc_buddy_huge_page(struct hstate
*h
,
1509 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1511 int order
= huge_page_order(h
);
1512 gfp_t gfp
= htlb_alloc_mask(h
)|__GFP_COMP
|__GFP_REPEAT
|__GFP_NOWARN
;
1513 unsigned int cpuset_mems_cookie
;
1516 * We need a VMA to get a memory policy. If we do not
1517 * have one, we use the 'nid' argument.
1519 * The mempolicy stuff below has some non-inlined bits
1520 * and calls ->vm_ops. That makes it hard to optimize at
1521 * compile-time, even when NUMA is off and it does
1522 * nothing. This helps the compiler optimize it out.
1524 if (!IS_ENABLED(CONFIG_NUMA
) || !vma
) {
1526 * If a specific node is requested, make sure to
1527 * get memory from there, but only when a node
1528 * is explicitly specified.
1530 if (nid
!= NUMA_NO_NODE
)
1531 gfp
|= __GFP_THISNODE
;
1533 * Make sure to call something that can handle
1536 return alloc_pages_node(nid
, gfp
, order
);
1540 * OK, so we have a VMA. Fetch the mempolicy and try to
1541 * allocate a huge page with it. We will only reach this
1542 * when CONFIG_NUMA=y.
1546 struct mempolicy
*mpol
;
1547 struct zonelist
*zl
;
1548 nodemask_t
*nodemask
;
1550 cpuset_mems_cookie
= read_mems_allowed_begin();
1551 zl
= huge_zonelist(vma
, addr
, gfp
, &mpol
, &nodemask
);
1552 mpol_cond_put(mpol
);
1553 page
= __alloc_pages_nodemask(gfp
, order
, zl
, nodemask
);
1556 } while (read_mems_allowed_retry(cpuset_mems_cookie
));
1562 * There are two ways to allocate a huge page:
1563 * 1. When you have a VMA and an address (like a fault)
1564 * 2. When you have no VMA (like when setting /proc/.../nr_hugepages)
1566 * 'vma' and 'addr' are only for (1). 'nid' is always NUMA_NO_NODE in
1567 * this case which signifies that the allocation should be done with
1568 * respect for the VMA's memory policy.
1570 * For (2), we ignore 'vma' and 'addr' and use 'nid' exclusively. This
1571 * implies that memory policies will not be taken in to account.
1573 static struct page
*__alloc_buddy_huge_page(struct hstate
*h
,
1574 struct vm_area_struct
*vma
, unsigned long addr
, int nid
)
1579 if (hstate_is_gigantic(h
))
1583 * Make sure that anyone specifying 'nid' is not also specifying a VMA.
1584 * This makes sure the caller is picking _one_ of the modes with which
1585 * we can call this function, not both.
1587 if (vma
|| (addr
!= -1)) {
1588 VM_WARN_ON_ONCE(addr
== -1);
1589 VM_WARN_ON_ONCE(nid
!= NUMA_NO_NODE
);
1592 * Assume we will successfully allocate the surplus page to
1593 * prevent racing processes from causing the surplus to exceed
1596 * This however introduces a different race, where a process B
1597 * tries to grow the static hugepage pool while alloc_pages() is
1598 * called by process A. B will only examine the per-node
1599 * counters in determining if surplus huge pages can be
1600 * converted to normal huge pages in adjust_pool_surplus(). A
1601 * won't be able to increment the per-node counter, until the
1602 * lock is dropped by B, but B doesn't drop hugetlb_lock until
1603 * no more huge pages can be converted from surplus to normal
1604 * state (and doesn't try to convert again). Thus, we have a
1605 * case where a surplus huge page exists, the pool is grown, and
1606 * the surplus huge page still exists after, even though it
1607 * should just have been converted to a normal huge page. This
1608 * does not leak memory, though, as the hugepage will be freed
1609 * once it is out of use. It also does not allow the counters to
1610 * go out of whack in adjust_pool_surplus() as we don't modify
1611 * the node values until we've gotten the hugepage and only the
1612 * per-node value is checked there.
1614 spin_lock(&hugetlb_lock
);
1615 if (h
->surplus_huge_pages
>= h
->nr_overcommit_huge_pages
) {
1616 spin_unlock(&hugetlb_lock
);
1620 h
->surplus_huge_pages
++;
1622 spin_unlock(&hugetlb_lock
);
1624 page
= __hugetlb_alloc_buddy_huge_page(h
, vma
, addr
, nid
);
1626 spin_lock(&hugetlb_lock
);
1628 INIT_LIST_HEAD(&page
->lru
);
1629 r_nid
= page_to_nid(page
);
1630 set_compound_page_dtor(page
, HUGETLB_PAGE_DTOR
);
1631 set_hugetlb_cgroup(page
, NULL
);
1633 * We incremented the global counters already
1635 h
->nr_huge_pages_node
[r_nid
]++;
1636 h
->surplus_huge_pages_node
[r_nid
]++;
1637 __count_vm_event(HTLB_BUDDY_PGALLOC
);
1640 h
->surplus_huge_pages
--;
1641 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL
);
1643 spin_unlock(&hugetlb_lock
);
1649 * Allocate a huge page from 'nid'. Note, 'nid' may be
1650 * NUMA_NO_NODE, which means that it may be allocated
1654 struct page
*__alloc_buddy_huge_page_no_mpol(struct hstate
*h
, int nid
)
1656 unsigned long addr
= -1;
1658 return __alloc_buddy_huge_page(h
, NULL
, addr
, nid
);
1662 * Use the VMA's mpolicy to allocate a huge page from the buddy.
1665 struct page
*__alloc_buddy_huge_page_with_mpol(struct hstate
*h
,
1666 struct vm_area_struct
*vma
, unsigned long addr
)
1668 return __alloc_buddy_huge_page(h
, vma
, addr
, NUMA_NO_NODE
);
1672 * This allocation function is useful in the context where vma is irrelevant.
1673 * E.g. soft-offlining uses this function because it only cares physical
1674 * address of error page.
1676 struct page
*alloc_huge_page_node(struct hstate
*h
, int nid
)
1678 struct page
*page
= NULL
;
1680 spin_lock(&hugetlb_lock
);
1681 if (h
->free_huge_pages
- h
->resv_huge_pages
> 0)
1682 page
= dequeue_huge_page_node(h
, nid
);
1683 spin_unlock(&hugetlb_lock
);
1686 page
= __alloc_buddy_huge_page_no_mpol(h
, nid
);
1692 * Increase the hugetlb pool such that it can accommodate a reservation
1695 static int gather_surplus_pages(struct hstate
*h
, int delta
)
1697 struct list_head surplus_list
;
1698 struct page
*page
, *tmp
;
1700 int needed
, allocated
;
1701 bool alloc_ok
= true;
1703 needed
= (h
->resv_huge_pages
+ delta
) - h
->free_huge_pages
;
1705 h
->resv_huge_pages
+= delta
;
1710 INIT_LIST_HEAD(&surplus_list
);
1714 spin_unlock(&hugetlb_lock
);
1715 for (i
= 0; i
< needed
; i
++) {
1716 page
= __alloc_buddy_huge_page_no_mpol(h
, NUMA_NO_NODE
);
1721 list_add(&page
->lru
, &surplus_list
);
1726 * After retaking hugetlb_lock, we need to recalculate 'needed'
1727 * because either resv_huge_pages or free_huge_pages may have changed.
1729 spin_lock(&hugetlb_lock
);
1730 needed
= (h
->resv_huge_pages
+ delta
) -
1731 (h
->free_huge_pages
+ allocated
);
1736 * We were not able to allocate enough pages to
1737 * satisfy the entire reservation so we free what
1738 * we've allocated so far.
1743 * The surplus_list now contains _at_least_ the number of extra pages
1744 * needed to accommodate the reservation. Add the appropriate number
1745 * of pages to the hugetlb pool and free the extras back to the buddy
1746 * allocator. Commit the entire reservation here to prevent another
1747 * process from stealing the pages as they are added to the pool but
1748 * before they are reserved.
1750 needed
+= allocated
;
1751 h
->resv_huge_pages
+= delta
;
1754 /* Free the needed pages to the hugetlb pool */
1755 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
) {
1759 * This page is now managed by the hugetlb allocator and has
1760 * no users -- drop the buddy allocator's reference.
1762 put_page_testzero(page
);
1763 VM_BUG_ON_PAGE(page_count(page
), page
);
1764 enqueue_huge_page(h
, page
);
1767 spin_unlock(&hugetlb_lock
);
1769 /* Free unnecessary surplus pages to the buddy allocator */
1770 list_for_each_entry_safe(page
, tmp
, &surplus_list
, lru
)
1772 spin_lock(&hugetlb_lock
);
1778 * This routine has two main purposes:
1779 * 1) Decrement the reservation count (resv_huge_pages) by the value passed
1780 * in unused_resv_pages. This corresponds to the prior adjustments made
1781 * to the associated reservation map.
1782 * 2) Free any unused surplus pages that may have been allocated to satisfy
1783 * the reservation. As many as unused_resv_pages may be freed.
1785 * Called with hugetlb_lock held. However, the lock could be dropped (and
1786 * reacquired) during calls to cond_resched_lock. Whenever dropping the lock,
1787 * we must make sure nobody else can claim pages we are in the process of
1788 * freeing. Do this by ensuring resv_huge_page always is greater than the
1789 * number of huge pages we plan to free when dropping the lock.
1791 static void return_unused_surplus_pages(struct hstate
*h
,
1792 unsigned long unused_resv_pages
)
1794 unsigned long nr_pages
;
1796 /* Cannot return gigantic pages currently */
1797 if (hstate_is_gigantic(h
))
1801 * Part (or even all) of the reservation could have been backed
1802 * by pre-allocated pages. Only free surplus pages.
1804 nr_pages
= min(unused_resv_pages
, h
->surplus_huge_pages
);
1807 * We want to release as many surplus pages as possible, spread
1808 * evenly across all nodes with memory. Iterate across these nodes
1809 * until we can no longer free unreserved surplus pages. This occurs
1810 * when the nodes with surplus pages have no free pages.
1811 * free_pool_huge_page() will balance the the freed pages across the
1812 * on-line nodes with memory and will handle the hstate accounting.
1814 * Note that we decrement resv_huge_pages as we free the pages. If
1815 * we drop the lock, resv_huge_pages will still be sufficiently large
1816 * to cover subsequent pages we may free.
1818 while (nr_pages
--) {
1819 h
->resv_huge_pages
--;
1820 unused_resv_pages
--;
1821 if (!free_pool_huge_page(h
, &node_states
[N_MEMORY
], 1))
1823 cond_resched_lock(&hugetlb_lock
);
1827 /* Fully uncommit the reservation */
1828 h
->resv_huge_pages
-= unused_resv_pages
;
1833 * vma_needs_reservation, vma_commit_reservation and vma_end_reservation
1834 * are used by the huge page allocation routines to manage reservations.
1836 * vma_needs_reservation is called to determine if the huge page at addr
1837 * within the vma has an associated reservation. If a reservation is
1838 * needed, the value 1 is returned. The caller is then responsible for
1839 * managing the global reservation and subpool usage counts. After
1840 * the huge page has been allocated, vma_commit_reservation is called
1841 * to add the page to the reservation map. If the page allocation fails,
1842 * the reservation must be ended instead of committed. vma_end_reservation
1843 * is called in such cases.
1845 * In the normal case, vma_commit_reservation returns the same value
1846 * as the preceding vma_needs_reservation call. The only time this
1847 * is not the case is if a reserve map was changed between calls. It
1848 * is the responsibility of the caller to notice the difference and
1849 * take appropriate action.
1851 * vma_add_reservation is used in error paths where a reservation must
1852 * be restored when a newly allocated huge page must be freed. It is
1853 * to be called after calling vma_needs_reservation to determine if a
1854 * reservation exists.
1856 enum vma_resv_mode
{
1862 static long __vma_reservation_common(struct hstate
*h
,
1863 struct vm_area_struct
*vma
, unsigned long addr
,
1864 enum vma_resv_mode mode
)
1866 struct resv_map
*resv
;
1870 resv
= vma_resv_map(vma
);
1874 idx
= vma_hugecache_offset(h
, vma
, addr
);
1876 case VMA_NEEDS_RESV
:
1877 ret
= region_chg(resv
, idx
, idx
+ 1);
1879 case VMA_COMMIT_RESV
:
1880 ret
= region_add(resv
, idx
, idx
+ 1);
1883 region_abort(resv
, idx
, idx
+ 1);
1887 if (vma
->vm_flags
& VM_MAYSHARE
)
1888 ret
= region_add(resv
, idx
, idx
+ 1);
1890 region_abort(resv
, idx
, idx
+ 1);
1891 ret
= region_del(resv
, idx
, idx
+ 1);
1898 if (vma
->vm_flags
& VM_MAYSHARE
)
1900 else if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) && ret
>= 0) {
1902 * In most cases, reserves always exist for private mappings.
1903 * However, a file associated with mapping could have been
1904 * hole punched or truncated after reserves were consumed.
1905 * As subsequent fault on such a range will not use reserves.
1906 * Subtle - The reserve map for private mappings has the
1907 * opposite meaning than that of shared mappings. If NO
1908 * entry is in the reserve map, it means a reservation exists.
1909 * If an entry exists in the reserve map, it means the
1910 * reservation has already been consumed. As a result, the
1911 * return value of this routine is the opposite of the
1912 * value returned from reserve map manipulation routines above.
1920 return ret
< 0 ? ret
: 0;
1923 static long vma_needs_reservation(struct hstate
*h
,
1924 struct vm_area_struct
*vma
, unsigned long addr
)
1926 return __vma_reservation_common(h
, vma
, addr
, VMA_NEEDS_RESV
);
1929 static long vma_commit_reservation(struct hstate
*h
,
1930 struct vm_area_struct
*vma
, unsigned long addr
)
1932 return __vma_reservation_common(h
, vma
, addr
, VMA_COMMIT_RESV
);
1935 static void vma_end_reservation(struct hstate
*h
,
1936 struct vm_area_struct
*vma
, unsigned long addr
)
1938 (void)__vma_reservation_common(h
, vma
, addr
, VMA_END_RESV
);
1941 static long vma_add_reservation(struct hstate
*h
,
1942 struct vm_area_struct
*vma
, unsigned long addr
)
1944 return __vma_reservation_common(h
, vma
, addr
, VMA_ADD_RESV
);
1948 * This routine is called to restore a reservation on error paths. In the
1949 * specific error paths, a huge page was allocated (via alloc_huge_page)
1950 * and is about to be freed. If a reservation for the page existed,
1951 * alloc_huge_page would have consumed the reservation and set PagePrivate
1952 * in the newly allocated page. When the page is freed via free_huge_page,
1953 * the global reservation count will be incremented if PagePrivate is set.
1954 * However, free_huge_page can not adjust the reserve map. Adjust the
1955 * reserve map here to be consistent with global reserve count adjustments
1956 * to be made by free_huge_page.
1958 static void restore_reserve_on_error(struct hstate
*h
,
1959 struct vm_area_struct
*vma
, unsigned long address
,
1962 if (unlikely(PagePrivate(page
))) {
1963 long rc
= vma_needs_reservation(h
, vma
, address
);
1965 if (unlikely(rc
< 0)) {
1967 * Rare out of memory condition in reserve map
1968 * manipulation. Clear PagePrivate so that
1969 * global reserve count will not be incremented
1970 * by free_huge_page. This will make it appear
1971 * as though the reservation for this page was
1972 * consumed. This may prevent the task from
1973 * faulting in the page at a later time. This
1974 * is better than inconsistent global huge page
1975 * accounting of reserve counts.
1977 ClearPagePrivate(page
);
1979 rc
= vma_add_reservation(h
, vma
, address
);
1980 if (unlikely(rc
< 0))
1982 * See above comment about rare out of
1985 ClearPagePrivate(page
);
1987 vma_end_reservation(h
, vma
, address
);
1991 struct page
*alloc_huge_page(struct vm_area_struct
*vma
,
1992 unsigned long addr
, int avoid_reserve
)
1994 struct hugepage_subpool
*spool
= subpool_vma(vma
);
1995 struct hstate
*h
= hstate_vma(vma
);
1997 long map_chg
, map_commit
;
2000 struct hugetlb_cgroup
*h_cg
;
2002 idx
= hstate_index(h
);
2004 * Examine the region/reserve map to determine if the process
2005 * has a reservation for the page to be allocated. A return
2006 * code of zero indicates a reservation exists (no change).
2008 map_chg
= gbl_chg
= vma_needs_reservation(h
, vma
, addr
);
2010 return ERR_PTR(-ENOMEM
);
2013 * Processes that did not create the mapping will have no
2014 * reserves as indicated by the region/reserve map. Check
2015 * that the allocation will not exceed the subpool limit.
2016 * Allocations for MAP_NORESERVE mappings also need to be
2017 * checked against any subpool limit.
2019 if (map_chg
|| avoid_reserve
) {
2020 gbl_chg
= hugepage_subpool_get_pages(spool
, 1);
2022 vma_end_reservation(h
, vma
, addr
);
2023 return ERR_PTR(-ENOSPC
);
2027 * Even though there was no reservation in the region/reserve
2028 * map, there could be reservations associated with the
2029 * subpool that can be used. This would be indicated if the
2030 * return value of hugepage_subpool_get_pages() is zero.
2031 * However, if avoid_reserve is specified we still avoid even
2032 * the subpool reservations.
2038 ret
= hugetlb_cgroup_charge_cgroup(idx
, pages_per_huge_page(h
), &h_cg
);
2040 goto out_subpool_put
;
2042 spin_lock(&hugetlb_lock
);
2044 * glb_chg is passed to indicate whether or not a page must be taken
2045 * from the global free pool (global change). gbl_chg == 0 indicates
2046 * a reservation exists for the allocation.
2048 page
= dequeue_huge_page_vma(h
, vma
, addr
, avoid_reserve
, gbl_chg
);
2050 spin_unlock(&hugetlb_lock
);
2051 page
= __alloc_buddy_huge_page_with_mpol(h
, vma
, addr
);
2053 goto out_uncharge_cgroup
;
2054 if (!avoid_reserve
&& vma_has_reserves(vma
, gbl_chg
)) {
2055 SetPagePrivate(page
);
2056 h
->resv_huge_pages
--;
2058 spin_lock(&hugetlb_lock
);
2059 list_move(&page
->lru
, &h
->hugepage_activelist
);
2062 hugetlb_cgroup_commit_charge(idx
, pages_per_huge_page(h
), h_cg
, page
);
2063 spin_unlock(&hugetlb_lock
);
2065 set_page_private(page
, (unsigned long)spool
);
2067 map_commit
= vma_commit_reservation(h
, vma
, addr
);
2068 if (unlikely(map_chg
> map_commit
)) {
2070 * The page was added to the reservation map between
2071 * vma_needs_reservation and vma_commit_reservation.
2072 * This indicates a race with hugetlb_reserve_pages.
2073 * Adjust for the subpool count incremented above AND
2074 * in hugetlb_reserve_pages for the same page. Also,
2075 * the reservation count added in hugetlb_reserve_pages
2076 * no longer applies.
2080 rsv_adjust
= hugepage_subpool_put_pages(spool
, 1);
2081 hugetlb_acct_memory(h
, -rsv_adjust
);
2085 out_uncharge_cgroup
:
2086 hugetlb_cgroup_uncharge_cgroup(idx
, pages_per_huge_page(h
), h_cg
);
2088 if (map_chg
|| avoid_reserve
)
2089 hugepage_subpool_put_pages(spool
, 1);
2090 vma_end_reservation(h
, vma
, addr
);
2091 return ERR_PTR(-ENOSPC
);
2095 * alloc_huge_page()'s wrapper which simply returns the page if allocation
2096 * succeeds, otherwise NULL. This function is called from new_vma_page(),
2097 * where no ERR_VALUE is expected to be returned.
2099 struct page
*alloc_huge_page_noerr(struct vm_area_struct
*vma
,
2100 unsigned long addr
, int avoid_reserve
)
2102 struct page
*page
= alloc_huge_page(vma
, addr
, avoid_reserve
);
2108 int __weak
alloc_bootmem_huge_page(struct hstate
*h
)
2110 struct huge_bootmem_page
*m
;
2113 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, &node_states
[N_MEMORY
]) {
2116 addr
= memblock_virt_alloc_try_nid_nopanic(
2117 huge_page_size(h
), huge_page_size(h
),
2118 0, BOOTMEM_ALLOC_ACCESSIBLE
, node
);
2121 * Use the beginning of the huge page to store the
2122 * huge_bootmem_page struct (until gather_bootmem
2123 * puts them into the mem_map).
2132 BUG_ON(!IS_ALIGNED(virt_to_phys(m
), huge_page_size(h
)));
2133 /* Put them into a private list first because mem_map is not up yet */
2134 list_add(&m
->list
, &huge_boot_pages
);
2139 static void __init
prep_compound_huge_page(struct page
*page
,
2142 if (unlikely(order
> (MAX_ORDER
- 1)))
2143 prep_compound_gigantic_page(page
, order
);
2145 prep_compound_page(page
, order
);
2148 /* Put bootmem huge pages into the standard lists after mem_map is up */
2149 static void __init
gather_bootmem_prealloc(void)
2151 struct huge_bootmem_page
*m
;
2153 list_for_each_entry(m
, &huge_boot_pages
, list
) {
2154 struct hstate
*h
= m
->hstate
;
2157 #ifdef CONFIG_HIGHMEM
2158 page
= pfn_to_page(m
->phys
>> PAGE_SHIFT
);
2159 memblock_free_late(__pa(m
),
2160 sizeof(struct huge_bootmem_page
));
2162 page
= virt_to_page(m
);
2164 WARN_ON(page_count(page
) != 1);
2165 prep_compound_huge_page(page
, h
->order
);
2166 WARN_ON(PageReserved(page
));
2167 prep_new_huge_page(h
, page
, page_to_nid(page
));
2169 * If we had gigantic hugepages allocated at boot time, we need
2170 * to restore the 'stolen' pages to totalram_pages in order to
2171 * fix confusing memory reports from free(1) and another
2172 * side-effects, like CommitLimit going negative.
2174 if (hstate_is_gigantic(h
))
2175 adjust_managed_page_count(page
, 1 << h
->order
);
2179 static void __init
hugetlb_hstate_alloc_pages(struct hstate
*h
)
2183 for (i
= 0; i
< h
->max_huge_pages
; ++i
) {
2184 if (hstate_is_gigantic(h
)) {
2185 if (!alloc_bootmem_huge_page(h
))
2187 } else if (!alloc_fresh_huge_page(h
,
2188 &node_states
[N_MEMORY
]))
2191 h
->max_huge_pages
= i
;
2194 static void __init
hugetlb_init_hstates(void)
2198 for_each_hstate(h
) {
2199 if (minimum_order
> huge_page_order(h
))
2200 minimum_order
= huge_page_order(h
);
2202 /* oversize hugepages were init'ed in early boot */
2203 if (!hstate_is_gigantic(h
))
2204 hugetlb_hstate_alloc_pages(h
);
2206 VM_BUG_ON(minimum_order
== UINT_MAX
);
2209 static char * __init
memfmt(char *buf
, unsigned long n
)
2211 if (n
>= (1UL << 30))
2212 sprintf(buf
, "%lu GB", n
>> 30);
2213 else if (n
>= (1UL << 20))
2214 sprintf(buf
, "%lu MB", n
>> 20);
2216 sprintf(buf
, "%lu KB", n
>> 10);
2220 static void __init
report_hugepages(void)
2224 for_each_hstate(h
) {
2226 pr_info("HugeTLB registered %s page size, pre-allocated %ld pages\n",
2227 memfmt(buf
, huge_page_size(h
)),
2228 h
->free_huge_pages
);
2232 #ifdef CONFIG_HIGHMEM
2233 static void try_to_free_low(struct hstate
*h
, unsigned long count
,
2234 nodemask_t
*nodes_allowed
)
2238 if (hstate_is_gigantic(h
))
2241 for_each_node_mask(i
, *nodes_allowed
) {
2242 struct page
*page
, *next
;
2243 struct list_head
*freel
= &h
->hugepage_freelists
[i
];
2244 list_for_each_entry_safe(page
, next
, freel
, lru
) {
2245 if (count
>= h
->nr_huge_pages
)
2247 if (PageHighMem(page
))
2249 list_del(&page
->lru
);
2250 update_and_free_page(h
, page
);
2251 h
->free_huge_pages
--;
2252 h
->free_huge_pages_node
[page_to_nid(page
)]--;
2257 static inline void try_to_free_low(struct hstate
*h
, unsigned long count
,
2258 nodemask_t
*nodes_allowed
)
2264 * Increment or decrement surplus_huge_pages. Keep node-specific counters
2265 * balanced by operating on them in a round-robin fashion.
2266 * Returns 1 if an adjustment was made.
2268 static int adjust_pool_surplus(struct hstate
*h
, nodemask_t
*nodes_allowed
,
2273 VM_BUG_ON(delta
!= -1 && delta
!= 1);
2276 for_each_node_mask_to_alloc(h
, nr_nodes
, node
, nodes_allowed
) {
2277 if (h
->surplus_huge_pages_node
[node
])
2281 for_each_node_mask_to_free(h
, nr_nodes
, node
, nodes_allowed
) {
2282 if (h
->surplus_huge_pages_node
[node
] <
2283 h
->nr_huge_pages_node
[node
])
2290 h
->surplus_huge_pages
+= delta
;
2291 h
->surplus_huge_pages_node
[node
] += delta
;
2295 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
2296 static unsigned long set_max_huge_pages(struct hstate
*h
, unsigned long count
,
2297 nodemask_t
*nodes_allowed
)
2299 unsigned long min_count
, ret
;
2301 if (hstate_is_gigantic(h
) && !gigantic_page_supported())
2302 return h
->max_huge_pages
;
2305 * Increase the pool size
2306 * First take pages out of surplus state. Then make up the
2307 * remaining difference by allocating fresh huge pages.
2309 * We might race with __alloc_buddy_huge_page() here and be unable
2310 * to convert a surplus huge page to a normal huge page. That is
2311 * not critical, though, it just means the overall size of the
2312 * pool might be one hugepage larger than it needs to be, but
2313 * within all the constraints specified by the sysctls.
2315 spin_lock(&hugetlb_lock
);
2316 while (h
->surplus_huge_pages
&& count
> persistent_huge_pages(h
)) {
2317 if (!adjust_pool_surplus(h
, nodes_allowed
, -1))
2321 while (count
> persistent_huge_pages(h
)) {
2323 * If this allocation races such that we no longer need the
2324 * page, free_huge_page will handle it by freeing the page
2325 * and reducing the surplus.
2327 spin_unlock(&hugetlb_lock
);
2329 /* yield cpu to avoid soft lockup */
2332 if (hstate_is_gigantic(h
))
2333 ret
= alloc_fresh_gigantic_page(h
, nodes_allowed
);
2335 ret
= alloc_fresh_huge_page(h
, nodes_allowed
);
2336 spin_lock(&hugetlb_lock
);
2340 /* Bail for signals. Probably ctrl-c from user */
2341 if (signal_pending(current
))
2346 * Decrease the pool size
2347 * First return free pages to the buddy allocator (being careful
2348 * to keep enough around to satisfy reservations). Then place
2349 * pages into surplus state as needed so the pool will shrink
2350 * to the desired size as pages become free.
2352 * By placing pages into the surplus state independent of the
2353 * overcommit value, we are allowing the surplus pool size to
2354 * exceed overcommit. There are few sane options here. Since
2355 * __alloc_buddy_huge_page() is checking the global counter,
2356 * though, we'll note that we're not allowed to exceed surplus
2357 * and won't grow the pool anywhere else. Not until one of the
2358 * sysctls are changed, or the surplus pages go out of use.
2360 min_count
= h
->resv_huge_pages
+ h
->nr_huge_pages
- h
->free_huge_pages
;
2361 min_count
= max(count
, min_count
);
2362 try_to_free_low(h
, min_count
, nodes_allowed
);
2363 while (min_count
< persistent_huge_pages(h
)) {
2364 if (!free_pool_huge_page(h
, nodes_allowed
, 0))
2366 cond_resched_lock(&hugetlb_lock
);
2368 while (count
< persistent_huge_pages(h
)) {
2369 if (!adjust_pool_surplus(h
, nodes_allowed
, 1))
2373 ret
= persistent_huge_pages(h
);
2374 spin_unlock(&hugetlb_lock
);
2378 #define HSTATE_ATTR_RO(_name) \
2379 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
2381 #define HSTATE_ATTR(_name) \
2382 static struct kobj_attribute _name##_attr = \
2383 __ATTR(_name, 0644, _name##_show, _name##_store)
2385 static struct kobject
*hugepages_kobj
;
2386 static struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2388 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
);
2390 static struct hstate
*kobj_to_hstate(struct kobject
*kobj
, int *nidp
)
2394 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2395 if (hstate_kobjs
[i
] == kobj
) {
2397 *nidp
= NUMA_NO_NODE
;
2401 return kobj_to_node_hstate(kobj
, nidp
);
2404 static ssize_t
nr_hugepages_show_common(struct kobject
*kobj
,
2405 struct kobj_attribute
*attr
, char *buf
)
2408 unsigned long nr_huge_pages
;
2411 h
= kobj_to_hstate(kobj
, &nid
);
2412 if (nid
== NUMA_NO_NODE
)
2413 nr_huge_pages
= h
->nr_huge_pages
;
2415 nr_huge_pages
= h
->nr_huge_pages_node
[nid
];
2417 return sprintf(buf
, "%lu\n", nr_huge_pages
);
2420 static ssize_t
__nr_hugepages_store_common(bool obey_mempolicy
,
2421 struct hstate
*h
, int nid
,
2422 unsigned long count
, size_t len
)
2425 NODEMASK_ALLOC(nodemask_t
, nodes_allowed
, GFP_KERNEL
| __GFP_NORETRY
);
2427 if (hstate_is_gigantic(h
) && !gigantic_page_supported()) {
2432 if (nid
== NUMA_NO_NODE
) {
2434 * global hstate attribute
2436 if (!(obey_mempolicy
&&
2437 init_nodemask_of_mempolicy(nodes_allowed
))) {
2438 NODEMASK_FREE(nodes_allowed
);
2439 nodes_allowed
= &node_states
[N_MEMORY
];
2441 } else if (nodes_allowed
) {
2443 * per node hstate attribute: adjust count to global,
2444 * but restrict alloc/free to the specified node.
2446 count
+= h
->nr_huge_pages
- h
->nr_huge_pages_node
[nid
];
2447 init_nodemask_of_node(nodes_allowed
, nid
);
2449 nodes_allowed
= &node_states
[N_MEMORY
];
2451 h
->max_huge_pages
= set_max_huge_pages(h
, count
, nodes_allowed
);
2453 if (nodes_allowed
!= &node_states
[N_MEMORY
])
2454 NODEMASK_FREE(nodes_allowed
);
2458 NODEMASK_FREE(nodes_allowed
);
2462 static ssize_t
nr_hugepages_store_common(bool obey_mempolicy
,
2463 struct kobject
*kobj
, const char *buf
,
2467 unsigned long count
;
2471 err
= kstrtoul(buf
, 10, &count
);
2475 h
= kobj_to_hstate(kobj
, &nid
);
2476 return __nr_hugepages_store_common(obey_mempolicy
, h
, nid
, count
, len
);
2479 static ssize_t
nr_hugepages_show(struct kobject
*kobj
,
2480 struct kobj_attribute
*attr
, char *buf
)
2482 return nr_hugepages_show_common(kobj
, attr
, buf
);
2485 static ssize_t
nr_hugepages_store(struct kobject
*kobj
,
2486 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2488 return nr_hugepages_store_common(false, kobj
, buf
, len
);
2490 HSTATE_ATTR(nr_hugepages
);
2495 * hstate attribute for optionally mempolicy-based constraint on persistent
2496 * huge page alloc/free.
2498 static ssize_t
nr_hugepages_mempolicy_show(struct kobject
*kobj
,
2499 struct kobj_attribute
*attr
, char *buf
)
2501 return nr_hugepages_show_common(kobj
, attr
, buf
);
2504 static ssize_t
nr_hugepages_mempolicy_store(struct kobject
*kobj
,
2505 struct kobj_attribute
*attr
, const char *buf
, size_t len
)
2507 return nr_hugepages_store_common(true, kobj
, buf
, len
);
2509 HSTATE_ATTR(nr_hugepages_mempolicy
);
2513 static ssize_t
nr_overcommit_hugepages_show(struct kobject
*kobj
,
2514 struct kobj_attribute
*attr
, char *buf
)
2516 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2517 return sprintf(buf
, "%lu\n", h
->nr_overcommit_huge_pages
);
2520 static ssize_t
nr_overcommit_hugepages_store(struct kobject
*kobj
,
2521 struct kobj_attribute
*attr
, const char *buf
, size_t count
)
2524 unsigned long input
;
2525 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2527 if (hstate_is_gigantic(h
))
2530 err
= kstrtoul(buf
, 10, &input
);
2534 spin_lock(&hugetlb_lock
);
2535 h
->nr_overcommit_huge_pages
= input
;
2536 spin_unlock(&hugetlb_lock
);
2540 HSTATE_ATTR(nr_overcommit_hugepages
);
2542 static ssize_t
free_hugepages_show(struct kobject
*kobj
,
2543 struct kobj_attribute
*attr
, char *buf
)
2546 unsigned long free_huge_pages
;
2549 h
= kobj_to_hstate(kobj
, &nid
);
2550 if (nid
== NUMA_NO_NODE
)
2551 free_huge_pages
= h
->free_huge_pages
;
2553 free_huge_pages
= h
->free_huge_pages_node
[nid
];
2555 return sprintf(buf
, "%lu\n", free_huge_pages
);
2557 HSTATE_ATTR_RO(free_hugepages
);
2559 static ssize_t
resv_hugepages_show(struct kobject
*kobj
,
2560 struct kobj_attribute
*attr
, char *buf
)
2562 struct hstate
*h
= kobj_to_hstate(kobj
, NULL
);
2563 return sprintf(buf
, "%lu\n", h
->resv_huge_pages
);
2565 HSTATE_ATTR_RO(resv_hugepages
);
2567 static ssize_t
surplus_hugepages_show(struct kobject
*kobj
,
2568 struct kobj_attribute
*attr
, char *buf
)
2571 unsigned long surplus_huge_pages
;
2574 h
= kobj_to_hstate(kobj
, &nid
);
2575 if (nid
== NUMA_NO_NODE
)
2576 surplus_huge_pages
= h
->surplus_huge_pages
;
2578 surplus_huge_pages
= h
->surplus_huge_pages_node
[nid
];
2580 return sprintf(buf
, "%lu\n", surplus_huge_pages
);
2582 HSTATE_ATTR_RO(surplus_hugepages
);
2584 static struct attribute
*hstate_attrs
[] = {
2585 &nr_hugepages_attr
.attr
,
2586 &nr_overcommit_hugepages_attr
.attr
,
2587 &free_hugepages_attr
.attr
,
2588 &resv_hugepages_attr
.attr
,
2589 &surplus_hugepages_attr
.attr
,
2591 &nr_hugepages_mempolicy_attr
.attr
,
2596 static struct attribute_group hstate_attr_group
= {
2597 .attrs
= hstate_attrs
,
2600 static int hugetlb_sysfs_add_hstate(struct hstate
*h
, struct kobject
*parent
,
2601 struct kobject
**hstate_kobjs
,
2602 struct attribute_group
*hstate_attr_group
)
2605 int hi
= hstate_index(h
);
2607 hstate_kobjs
[hi
] = kobject_create_and_add(h
->name
, parent
);
2608 if (!hstate_kobjs
[hi
])
2611 retval
= sysfs_create_group(hstate_kobjs
[hi
], hstate_attr_group
);
2613 kobject_put(hstate_kobjs
[hi
]);
2618 static void __init
hugetlb_sysfs_init(void)
2623 hugepages_kobj
= kobject_create_and_add("hugepages", mm_kobj
);
2624 if (!hugepages_kobj
)
2627 for_each_hstate(h
) {
2628 err
= hugetlb_sysfs_add_hstate(h
, hugepages_kobj
,
2629 hstate_kobjs
, &hstate_attr_group
);
2631 pr_err("Hugetlb: Unable to add hstate %s", h
->name
);
2638 * node_hstate/s - associate per node hstate attributes, via their kobjects,
2639 * with node devices in node_devices[] using a parallel array. The array
2640 * index of a node device or _hstate == node id.
2641 * This is here to avoid any static dependency of the node device driver, in
2642 * the base kernel, on the hugetlb module.
2644 struct node_hstate
{
2645 struct kobject
*hugepages_kobj
;
2646 struct kobject
*hstate_kobjs
[HUGE_MAX_HSTATE
];
2648 static struct node_hstate node_hstates
[MAX_NUMNODES
];
2651 * A subset of global hstate attributes for node devices
2653 static struct attribute
*per_node_hstate_attrs
[] = {
2654 &nr_hugepages_attr
.attr
,
2655 &free_hugepages_attr
.attr
,
2656 &surplus_hugepages_attr
.attr
,
2660 static struct attribute_group per_node_hstate_attr_group
= {
2661 .attrs
= per_node_hstate_attrs
,
2665 * kobj_to_node_hstate - lookup global hstate for node device hstate attr kobj.
2666 * Returns node id via non-NULL nidp.
2668 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2672 for (nid
= 0; nid
< nr_node_ids
; nid
++) {
2673 struct node_hstate
*nhs
= &node_hstates
[nid
];
2675 for (i
= 0; i
< HUGE_MAX_HSTATE
; i
++)
2676 if (nhs
->hstate_kobjs
[i
] == kobj
) {
2688 * Unregister hstate attributes from a single node device.
2689 * No-op if no hstate attributes attached.
2691 static void hugetlb_unregister_node(struct node
*node
)
2694 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2696 if (!nhs
->hugepages_kobj
)
2697 return; /* no hstate attributes */
2699 for_each_hstate(h
) {
2700 int idx
= hstate_index(h
);
2701 if (nhs
->hstate_kobjs
[idx
]) {
2702 kobject_put(nhs
->hstate_kobjs
[idx
]);
2703 nhs
->hstate_kobjs
[idx
] = NULL
;
2707 kobject_put(nhs
->hugepages_kobj
);
2708 nhs
->hugepages_kobj
= NULL
;
2713 * Register hstate attributes for a single node device.
2714 * No-op if attributes already registered.
2716 static void hugetlb_register_node(struct node
*node
)
2719 struct node_hstate
*nhs
= &node_hstates
[node
->dev
.id
];
2722 if (nhs
->hugepages_kobj
)
2723 return; /* already allocated */
2725 nhs
->hugepages_kobj
= kobject_create_and_add("hugepages",
2727 if (!nhs
->hugepages_kobj
)
2730 for_each_hstate(h
) {
2731 err
= hugetlb_sysfs_add_hstate(h
, nhs
->hugepages_kobj
,
2733 &per_node_hstate_attr_group
);
2735 pr_err("Hugetlb: Unable to add hstate %s for node %d\n",
2736 h
->name
, node
->dev
.id
);
2737 hugetlb_unregister_node(node
);
2744 * hugetlb init time: register hstate attributes for all registered node
2745 * devices of nodes that have memory. All on-line nodes should have
2746 * registered their associated device by this time.
2748 static void __init
hugetlb_register_all_nodes(void)
2752 for_each_node_state(nid
, N_MEMORY
) {
2753 struct node
*node
= node_devices
[nid
];
2754 if (node
->dev
.id
== nid
)
2755 hugetlb_register_node(node
);
2759 * Let the node device driver know we're here so it can
2760 * [un]register hstate attributes on node hotplug.
2762 register_hugetlbfs_with_node(hugetlb_register_node
,
2763 hugetlb_unregister_node
);
2765 #else /* !CONFIG_NUMA */
2767 static struct hstate
*kobj_to_node_hstate(struct kobject
*kobj
, int *nidp
)
2775 static void hugetlb_register_all_nodes(void) { }
2779 static int __init
hugetlb_init(void)
2783 if (!hugepages_supported())
2786 if (!size_to_hstate(default_hstate_size
)) {
2787 default_hstate_size
= HPAGE_SIZE
;
2788 if (!size_to_hstate(default_hstate_size
))
2789 hugetlb_add_hstate(HUGETLB_PAGE_ORDER
);
2791 default_hstate_idx
= hstate_index(size_to_hstate(default_hstate_size
));
2792 if (default_hstate_max_huge_pages
) {
2793 if (!default_hstate
.max_huge_pages
)
2794 default_hstate
.max_huge_pages
= default_hstate_max_huge_pages
;
2797 hugetlb_init_hstates();
2798 gather_bootmem_prealloc();
2801 hugetlb_sysfs_init();
2802 hugetlb_register_all_nodes();
2803 hugetlb_cgroup_file_init();
2806 num_fault_mutexes
= roundup_pow_of_two(8 * num_possible_cpus());
2808 num_fault_mutexes
= 1;
2810 hugetlb_fault_mutex_table
=
2811 kmalloc(sizeof(struct mutex
) * num_fault_mutexes
, GFP_KERNEL
);
2812 BUG_ON(!hugetlb_fault_mutex_table
);
2814 for (i
= 0; i
< num_fault_mutexes
; i
++)
2815 mutex_init(&hugetlb_fault_mutex_table
[i
]);
2818 subsys_initcall(hugetlb_init
);
2820 /* Should be called on processing a hugepagesz=... option */
2821 void __init
hugetlb_bad_size(void)
2823 parsed_valid_hugepagesz
= false;
2826 void __init
hugetlb_add_hstate(unsigned int order
)
2831 if (size_to_hstate(PAGE_SIZE
<< order
)) {
2832 pr_warn("hugepagesz= specified twice, ignoring\n");
2835 BUG_ON(hugetlb_max_hstate
>= HUGE_MAX_HSTATE
);
2837 h
= &hstates
[hugetlb_max_hstate
++];
2839 h
->mask
= ~((1ULL << (order
+ PAGE_SHIFT
)) - 1);
2840 h
->nr_huge_pages
= 0;
2841 h
->free_huge_pages
= 0;
2842 for (i
= 0; i
< MAX_NUMNODES
; ++i
)
2843 INIT_LIST_HEAD(&h
->hugepage_freelists
[i
]);
2844 INIT_LIST_HEAD(&h
->hugepage_activelist
);
2845 h
->next_nid_to_alloc
= first_memory_node
;
2846 h
->next_nid_to_free
= first_memory_node
;
2847 snprintf(h
->name
, HSTATE_NAME_LEN
, "hugepages-%lukB",
2848 huge_page_size(h
)/1024);
2853 static int __init
hugetlb_nrpages_setup(char *s
)
2856 static unsigned long *last_mhp
;
2858 if (!parsed_valid_hugepagesz
) {
2859 pr_warn("hugepages = %s preceded by "
2860 "an unsupported hugepagesz, ignoring\n", s
);
2861 parsed_valid_hugepagesz
= true;
2865 * !hugetlb_max_hstate means we haven't parsed a hugepagesz= parameter yet,
2866 * so this hugepages= parameter goes to the "default hstate".
2868 else if (!hugetlb_max_hstate
)
2869 mhp
= &default_hstate_max_huge_pages
;
2871 mhp
= &parsed_hstate
->max_huge_pages
;
2873 if (mhp
== last_mhp
) {
2874 pr_warn("hugepages= specified twice without interleaving hugepagesz=, ignoring\n");
2878 if (sscanf(s
, "%lu", mhp
) <= 0)
2882 * Global state is always initialized later in hugetlb_init.
2883 * But we need to allocate >= MAX_ORDER hstates here early to still
2884 * use the bootmem allocator.
2886 if (hugetlb_max_hstate
&& parsed_hstate
->order
>= MAX_ORDER
)
2887 hugetlb_hstate_alloc_pages(parsed_hstate
);
2893 __setup("hugepages=", hugetlb_nrpages_setup
);
2895 static int __init
hugetlb_default_setup(char *s
)
2897 default_hstate_size
= memparse(s
, &s
);
2900 __setup("default_hugepagesz=", hugetlb_default_setup
);
2902 static unsigned int cpuset_mems_nr(unsigned int *array
)
2905 unsigned int nr
= 0;
2907 for_each_node_mask(node
, cpuset_current_mems_allowed
)
2913 #ifdef CONFIG_SYSCTL
2914 static int hugetlb_sysctl_handler_common(bool obey_mempolicy
,
2915 struct ctl_table
*table
, int write
,
2916 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2918 struct hstate
*h
= &default_hstate
;
2919 unsigned long tmp
= h
->max_huge_pages
;
2922 if (!hugepages_supported())
2926 table
->maxlen
= sizeof(unsigned long);
2927 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2932 ret
= __nr_hugepages_store_common(obey_mempolicy
, h
,
2933 NUMA_NO_NODE
, tmp
, *length
);
2938 int hugetlb_sysctl_handler(struct ctl_table
*table
, int write
,
2939 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2942 return hugetlb_sysctl_handler_common(false, table
, write
,
2943 buffer
, length
, ppos
);
2947 int hugetlb_mempolicy_sysctl_handler(struct ctl_table
*table
, int write
,
2948 void __user
*buffer
, size_t *length
, loff_t
*ppos
)
2950 return hugetlb_sysctl_handler_common(true, table
, write
,
2951 buffer
, length
, ppos
);
2953 #endif /* CONFIG_NUMA */
2955 int hugetlb_overcommit_handler(struct ctl_table
*table
, int write
,
2956 void __user
*buffer
,
2957 size_t *length
, loff_t
*ppos
)
2959 struct hstate
*h
= &default_hstate
;
2963 if (!hugepages_supported())
2966 tmp
= h
->nr_overcommit_huge_pages
;
2968 if (write
&& hstate_is_gigantic(h
))
2972 table
->maxlen
= sizeof(unsigned long);
2973 ret
= proc_doulongvec_minmax(table
, write
, buffer
, length
, ppos
);
2978 spin_lock(&hugetlb_lock
);
2979 h
->nr_overcommit_huge_pages
= tmp
;
2980 spin_unlock(&hugetlb_lock
);
2986 #endif /* CONFIG_SYSCTL */
2988 void hugetlb_report_meminfo(struct seq_file
*m
)
2990 struct hstate
*h
= &default_hstate
;
2991 if (!hugepages_supported())
2994 "HugePages_Total: %5lu\n"
2995 "HugePages_Free: %5lu\n"
2996 "HugePages_Rsvd: %5lu\n"
2997 "HugePages_Surp: %5lu\n"
2998 "Hugepagesize: %8lu kB\n",
3002 h
->surplus_huge_pages
,
3003 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3006 int hugetlb_report_node_meminfo(int nid
, char *buf
)
3008 struct hstate
*h
= &default_hstate
;
3009 if (!hugepages_supported())
3012 "Node %d HugePages_Total: %5u\n"
3013 "Node %d HugePages_Free: %5u\n"
3014 "Node %d HugePages_Surp: %5u\n",
3015 nid
, h
->nr_huge_pages_node
[nid
],
3016 nid
, h
->free_huge_pages_node
[nid
],
3017 nid
, h
->surplus_huge_pages_node
[nid
]);
3020 void hugetlb_show_meminfo(void)
3025 if (!hugepages_supported())
3028 for_each_node_state(nid
, N_MEMORY
)
3030 pr_info("Node %d hugepages_total=%u hugepages_free=%u hugepages_surp=%u hugepages_size=%lukB\n",
3032 h
->nr_huge_pages_node
[nid
],
3033 h
->free_huge_pages_node
[nid
],
3034 h
->surplus_huge_pages_node
[nid
],
3035 1UL << (huge_page_order(h
) + PAGE_SHIFT
- 10));
3038 void hugetlb_report_usage(struct seq_file
*m
, struct mm_struct
*mm
)
3040 seq_printf(m
, "HugetlbPages:\t%8lu kB\n",
3041 atomic_long_read(&mm
->hugetlb_usage
) << (PAGE_SHIFT
- 10));
3044 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
3045 unsigned long hugetlb_total_pages(void)
3048 unsigned long nr_total_pages
= 0;
3051 nr_total_pages
+= h
->nr_huge_pages
* pages_per_huge_page(h
);
3052 return nr_total_pages
;
3055 static int hugetlb_acct_memory(struct hstate
*h
, long delta
)
3059 spin_lock(&hugetlb_lock
);
3061 * When cpuset is configured, it breaks the strict hugetlb page
3062 * reservation as the accounting is done on a global variable. Such
3063 * reservation is completely rubbish in the presence of cpuset because
3064 * the reservation is not checked against page availability for the
3065 * current cpuset. Application can still potentially OOM'ed by kernel
3066 * with lack of free htlb page in cpuset that the task is in.
3067 * Attempt to enforce strict accounting with cpuset is almost
3068 * impossible (or too ugly) because cpuset is too fluid that
3069 * task or memory node can be dynamically moved between cpusets.
3071 * The change of semantics for shared hugetlb mapping with cpuset is
3072 * undesirable. However, in order to preserve some of the semantics,
3073 * we fall back to check against current free page availability as
3074 * a best attempt and hopefully to minimize the impact of changing
3075 * semantics that cpuset has.
3078 if (gather_surplus_pages(h
, delta
) < 0)
3081 if (delta
> cpuset_mems_nr(h
->free_huge_pages_node
)) {
3082 return_unused_surplus_pages(h
, delta
);
3089 return_unused_surplus_pages(h
, (unsigned long) -delta
);
3092 spin_unlock(&hugetlb_lock
);
3096 static void hugetlb_vm_op_open(struct vm_area_struct
*vma
)
3098 struct resv_map
*resv
= vma_resv_map(vma
);
3101 * This new VMA should share its siblings reservation map if present.
3102 * The VMA will only ever have a valid reservation map pointer where
3103 * it is being copied for another still existing VMA. As that VMA
3104 * has a reference to the reservation map it cannot disappear until
3105 * after this open call completes. It is therefore safe to take a
3106 * new reference here without additional locking.
3108 if (resv
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3109 kref_get(&resv
->refs
);
3112 static void hugetlb_vm_op_close(struct vm_area_struct
*vma
)
3114 struct hstate
*h
= hstate_vma(vma
);
3115 struct resv_map
*resv
= vma_resv_map(vma
);
3116 struct hugepage_subpool
*spool
= subpool_vma(vma
);
3117 unsigned long reserve
, start
, end
;
3120 if (!resv
|| !is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
3123 start
= vma_hugecache_offset(h
, vma
, vma
->vm_start
);
3124 end
= vma_hugecache_offset(h
, vma
, vma
->vm_end
);
3126 reserve
= (end
- start
) - region_count(resv
, start
, end
);
3128 kref_put(&resv
->refs
, resv_map_release
);
3132 * Decrement reserve counts. The global reserve count may be
3133 * adjusted if the subpool has a minimum size.
3135 gbl_reserve
= hugepage_subpool_put_pages(spool
, reserve
);
3136 hugetlb_acct_memory(h
, -gbl_reserve
);
3141 * We cannot handle pagefaults against hugetlb pages at all. They cause
3142 * handle_mm_fault() to try to instantiate regular-sized pages in the
3143 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
3146 static int hugetlb_vm_op_fault(struct vm_fault
*vmf
)
3152 const struct vm_operations_struct hugetlb_vm_ops
= {
3153 .fault
= hugetlb_vm_op_fault
,
3154 .open
= hugetlb_vm_op_open
,
3155 .close
= hugetlb_vm_op_close
,
3158 static pte_t
make_huge_pte(struct vm_area_struct
*vma
, struct page
*page
,
3164 entry
= huge_pte_mkwrite(huge_pte_mkdirty(mk_huge_pte(page
,
3165 vma
->vm_page_prot
)));
3167 entry
= huge_pte_wrprotect(mk_huge_pte(page
,
3168 vma
->vm_page_prot
));
3170 entry
= pte_mkyoung(entry
);
3171 entry
= pte_mkhuge(entry
);
3172 entry
= arch_make_huge_pte(entry
, vma
, page
, writable
);
3177 static void set_huge_ptep_writable(struct vm_area_struct
*vma
,
3178 unsigned long address
, pte_t
*ptep
)
3182 entry
= huge_pte_mkwrite(huge_pte_mkdirty(huge_ptep_get(ptep
)));
3183 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
, 1))
3184 update_mmu_cache(vma
, address
, ptep
);
3187 static int is_hugetlb_entry_migration(pte_t pte
)
3191 if (huge_pte_none(pte
) || pte_present(pte
))
3193 swp
= pte_to_swp_entry(pte
);
3194 if (non_swap_entry(swp
) && is_migration_entry(swp
))
3200 static int is_hugetlb_entry_hwpoisoned(pte_t pte
)
3204 if (huge_pte_none(pte
) || pte_present(pte
))
3206 swp
= pte_to_swp_entry(pte
);
3207 if (non_swap_entry(swp
) && is_hwpoison_entry(swp
))
3213 int copy_hugetlb_page_range(struct mm_struct
*dst
, struct mm_struct
*src
,
3214 struct vm_area_struct
*vma
)
3216 pte_t
*src_pte
, *dst_pte
, entry
;
3217 struct page
*ptepage
;
3220 struct hstate
*h
= hstate_vma(vma
);
3221 unsigned long sz
= huge_page_size(h
);
3222 unsigned long mmun_start
; /* For mmu_notifiers */
3223 unsigned long mmun_end
; /* For mmu_notifiers */
3226 cow
= (vma
->vm_flags
& (VM_SHARED
| VM_MAYWRITE
)) == VM_MAYWRITE
;
3228 mmun_start
= vma
->vm_start
;
3229 mmun_end
= vma
->vm_end
;
3231 mmu_notifier_invalidate_range_start(src
, mmun_start
, mmun_end
);
3233 for (addr
= vma
->vm_start
; addr
< vma
->vm_end
; addr
+= sz
) {
3234 spinlock_t
*src_ptl
, *dst_ptl
;
3235 src_pte
= huge_pte_offset(src
, addr
);
3238 dst_pte
= huge_pte_alloc(dst
, addr
, sz
);
3244 /* If the pagetables are shared don't copy or take references */
3245 if (dst_pte
== src_pte
)
3248 dst_ptl
= huge_pte_lock(h
, dst
, dst_pte
);
3249 src_ptl
= huge_pte_lockptr(h
, src
, src_pte
);
3250 spin_lock_nested(src_ptl
, SINGLE_DEPTH_NESTING
);
3251 entry
= huge_ptep_get(src_pte
);
3252 if (huge_pte_none(entry
)) { /* skip none entry */
3254 } else if (unlikely(is_hugetlb_entry_migration(entry
) ||
3255 is_hugetlb_entry_hwpoisoned(entry
))) {
3256 swp_entry_t swp_entry
= pte_to_swp_entry(entry
);
3258 if (is_write_migration_entry(swp_entry
) && cow
) {
3260 * COW mappings require pages in both
3261 * parent and child to be set to read.
3263 make_migration_entry_read(&swp_entry
);
3264 entry
= swp_entry_to_pte(swp_entry
);
3265 set_huge_pte_at(src
, addr
, src_pte
, entry
);
3267 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3270 huge_ptep_set_wrprotect(src
, addr
, src_pte
);
3271 mmu_notifier_invalidate_range(src
, mmun_start
,
3274 entry
= huge_ptep_get(src_pte
);
3275 ptepage
= pte_page(entry
);
3277 page_dup_rmap(ptepage
, true);
3278 set_huge_pte_at(dst
, addr
, dst_pte
, entry
);
3279 hugetlb_count_add(pages_per_huge_page(h
), dst
);
3281 spin_unlock(src_ptl
);
3282 spin_unlock(dst_ptl
);
3286 mmu_notifier_invalidate_range_end(src
, mmun_start
, mmun_end
);
3291 void __unmap_hugepage_range(struct mmu_gather
*tlb
, struct vm_area_struct
*vma
,
3292 unsigned long start
, unsigned long end
,
3293 struct page
*ref_page
)
3295 struct mm_struct
*mm
= vma
->vm_mm
;
3296 unsigned long address
;
3301 struct hstate
*h
= hstate_vma(vma
);
3302 unsigned long sz
= huge_page_size(h
);
3303 const unsigned long mmun_start
= start
; /* For mmu_notifiers */
3304 const unsigned long mmun_end
= end
; /* For mmu_notifiers */
3306 WARN_ON(!is_vm_hugetlb_page(vma
));
3307 BUG_ON(start
& ~huge_page_mask(h
));
3308 BUG_ON(end
& ~huge_page_mask(h
));
3311 * This is a hugetlb vma, all the pte entries should point
3314 tlb_remove_check_page_size_change(tlb
, sz
);
3315 tlb_start_vma(tlb
, vma
);
3316 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3318 for (; address
< end
; address
+= sz
) {
3319 ptep
= huge_pte_offset(mm
, address
);
3323 ptl
= huge_pte_lock(h
, mm
, ptep
);
3324 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
3329 pte
= huge_ptep_get(ptep
);
3330 if (huge_pte_none(pte
)) {
3336 * Migrating hugepage or HWPoisoned hugepage is already
3337 * unmapped and its refcount is dropped, so just clear pte here.
3339 if (unlikely(!pte_present(pte
))) {
3340 huge_pte_clear(mm
, address
, ptep
);
3345 page
= pte_page(pte
);
3347 * If a reference page is supplied, it is because a specific
3348 * page is being unmapped, not a range. Ensure the page we
3349 * are about to unmap is the actual page of interest.
3352 if (page
!= ref_page
) {
3357 * Mark the VMA as having unmapped its page so that
3358 * future faults in this VMA will fail rather than
3359 * looking like data was lost
3361 set_vma_resv_flags(vma
, HPAGE_RESV_UNMAPPED
);
3364 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
3365 tlb_remove_huge_tlb_entry(h
, tlb
, ptep
, address
);
3366 if (huge_pte_dirty(pte
))
3367 set_page_dirty(page
);
3369 hugetlb_count_sub(pages_per_huge_page(h
), mm
);
3370 page_remove_rmap(page
, true);
3373 tlb_remove_page_size(tlb
, page
, huge_page_size(h
));
3375 * Bail out after unmapping reference page if supplied
3380 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3381 tlb_end_vma(tlb
, vma
);
3384 void __unmap_hugepage_range_final(struct mmu_gather
*tlb
,
3385 struct vm_area_struct
*vma
, unsigned long start
,
3386 unsigned long end
, struct page
*ref_page
)
3388 __unmap_hugepage_range(tlb
, vma
, start
, end
, ref_page
);
3391 * Clear this flag so that x86's huge_pmd_share page_table_shareable
3392 * test will fail on a vma being torn down, and not grab a page table
3393 * on its way out. We're lucky that the flag has such an appropriate
3394 * name, and can in fact be safely cleared here. We could clear it
3395 * before the __unmap_hugepage_range above, but all that's necessary
3396 * is to clear it before releasing the i_mmap_rwsem. This works
3397 * because in the context this is called, the VMA is about to be
3398 * destroyed and the i_mmap_rwsem is held.
3400 vma
->vm_flags
&= ~VM_MAYSHARE
;
3403 void unmap_hugepage_range(struct vm_area_struct
*vma
, unsigned long start
,
3404 unsigned long end
, struct page
*ref_page
)
3406 struct mm_struct
*mm
;
3407 struct mmu_gather tlb
;
3411 tlb_gather_mmu(&tlb
, mm
, start
, end
);
3412 __unmap_hugepage_range(&tlb
, vma
, start
, end
, ref_page
);
3413 tlb_finish_mmu(&tlb
, start
, end
);
3417 * This is called when the original mapper is failing to COW a MAP_PRIVATE
3418 * mappping it owns the reserve page for. The intention is to unmap the page
3419 * from other VMAs and let the children be SIGKILLed if they are faulting the
3422 static void unmap_ref_private(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3423 struct page
*page
, unsigned long address
)
3425 struct hstate
*h
= hstate_vma(vma
);
3426 struct vm_area_struct
*iter_vma
;
3427 struct address_space
*mapping
;
3431 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
3432 * from page cache lookup which is in HPAGE_SIZE units.
3434 address
= address
& huge_page_mask(h
);
3435 pgoff
= ((address
- vma
->vm_start
) >> PAGE_SHIFT
) +
3437 mapping
= vma
->vm_file
->f_mapping
;
3440 * Take the mapping lock for the duration of the table walk. As
3441 * this mapping should be shared between all the VMAs,
3442 * __unmap_hugepage_range() is called as the lock is already held
3444 i_mmap_lock_write(mapping
);
3445 vma_interval_tree_foreach(iter_vma
, &mapping
->i_mmap
, pgoff
, pgoff
) {
3446 /* Do not unmap the current VMA */
3447 if (iter_vma
== vma
)
3451 * Shared VMAs have their own reserves and do not affect
3452 * MAP_PRIVATE accounting but it is possible that a shared
3453 * VMA is using the same page so check and skip such VMAs.
3455 if (iter_vma
->vm_flags
& VM_MAYSHARE
)
3459 * Unmap the page from other VMAs without their own reserves.
3460 * They get marked to be SIGKILLed if they fault in these
3461 * areas. This is because a future no-page fault on this VMA
3462 * could insert a zeroed page instead of the data existing
3463 * from the time of fork. This would look like data corruption
3465 if (!is_vma_resv_set(iter_vma
, HPAGE_RESV_OWNER
))
3466 unmap_hugepage_range(iter_vma
, address
,
3467 address
+ huge_page_size(h
), page
);
3469 i_mmap_unlock_write(mapping
);
3473 * Hugetlb_cow() should be called with page lock of the original hugepage held.
3474 * Called with hugetlb_instantiation_mutex held and pte_page locked so we
3475 * cannot race with other handlers or page migration.
3476 * Keep the pte_same checks anyway to make transition from the mutex easier.
3478 static int hugetlb_cow(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3479 unsigned long address
, pte_t
*ptep
,
3480 struct page
*pagecache_page
, spinlock_t
*ptl
)
3483 struct hstate
*h
= hstate_vma(vma
);
3484 struct page
*old_page
, *new_page
;
3485 int ret
= 0, outside_reserve
= 0;
3486 unsigned long mmun_start
; /* For mmu_notifiers */
3487 unsigned long mmun_end
; /* For mmu_notifiers */
3489 pte
= huge_ptep_get(ptep
);
3490 old_page
= pte_page(pte
);
3493 /* If no-one else is actually using this page, avoid the copy
3494 * and just make the page writable */
3495 if (page_mapcount(old_page
) == 1 && PageAnon(old_page
)) {
3496 page_move_anon_rmap(old_page
, vma
);
3497 set_huge_ptep_writable(vma
, address
, ptep
);
3502 * If the process that created a MAP_PRIVATE mapping is about to
3503 * perform a COW due to a shared page count, attempt to satisfy
3504 * the allocation without using the existing reserves. The pagecache
3505 * page is used to determine if the reserve at this address was
3506 * consumed or not. If reserves were used, a partial faulted mapping
3507 * at the time of fork() could consume its reserves on COW instead
3508 * of the full address range.
3510 if (is_vma_resv_set(vma
, HPAGE_RESV_OWNER
) &&
3511 old_page
!= pagecache_page
)
3512 outside_reserve
= 1;
3517 * Drop page table lock as buddy allocator may be called. It will
3518 * be acquired again before returning to the caller, as expected.
3521 new_page
= alloc_huge_page(vma
, address
, outside_reserve
);
3523 if (IS_ERR(new_page
)) {
3525 * If a process owning a MAP_PRIVATE mapping fails to COW,
3526 * it is due to references held by a child and an insufficient
3527 * huge page pool. To guarantee the original mappers
3528 * reliability, unmap the page from child processes. The child
3529 * may get SIGKILLed if it later faults.
3531 if (outside_reserve
) {
3533 BUG_ON(huge_pte_none(pte
));
3534 unmap_ref_private(mm
, vma
, old_page
, address
);
3535 BUG_ON(huge_pte_none(pte
));
3537 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3539 pte_same(huge_ptep_get(ptep
), pte
)))
3540 goto retry_avoidcopy
;
3542 * race occurs while re-acquiring page table
3543 * lock, and our job is done.
3548 ret
= (PTR_ERR(new_page
) == -ENOMEM
) ?
3549 VM_FAULT_OOM
: VM_FAULT_SIGBUS
;
3550 goto out_release_old
;
3554 * When the original hugepage is shared one, it does not have
3555 * anon_vma prepared.
3557 if (unlikely(anon_vma_prepare(vma
))) {
3559 goto out_release_all
;
3562 copy_user_huge_page(new_page
, old_page
, address
, vma
,
3563 pages_per_huge_page(h
));
3564 __SetPageUptodate(new_page
);
3565 set_page_huge_active(new_page
);
3567 mmun_start
= address
& huge_page_mask(h
);
3568 mmun_end
= mmun_start
+ huge_page_size(h
);
3569 mmu_notifier_invalidate_range_start(mm
, mmun_start
, mmun_end
);
3572 * Retake the page table lock to check for racing updates
3573 * before the page tables are altered
3576 ptep
= huge_pte_offset(mm
, address
& huge_page_mask(h
));
3577 if (likely(ptep
&& pte_same(huge_ptep_get(ptep
), pte
))) {
3578 ClearPagePrivate(new_page
);
3581 huge_ptep_clear_flush(vma
, address
, ptep
);
3582 mmu_notifier_invalidate_range(mm
, mmun_start
, mmun_end
);
3583 set_huge_pte_at(mm
, address
, ptep
,
3584 make_huge_pte(vma
, new_page
, 1));
3585 page_remove_rmap(old_page
, true);
3586 hugepage_add_new_anon_rmap(new_page
, vma
, address
);
3587 /* Make the old page be freed below */
3588 new_page
= old_page
;
3591 mmu_notifier_invalidate_range_end(mm
, mmun_start
, mmun_end
);
3593 restore_reserve_on_error(h
, vma
, address
, new_page
);
3598 spin_lock(ptl
); /* Caller expects lock to be held */
3602 /* Return the pagecache page at a given address within a VMA */
3603 static struct page
*hugetlbfs_pagecache_page(struct hstate
*h
,
3604 struct vm_area_struct
*vma
, unsigned long address
)
3606 struct address_space
*mapping
;
3609 mapping
= vma
->vm_file
->f_mapping
;
3610 idx
= vma_hugecache_offset(h
, vma
, address
);
3612 return find_lock_page(mapping
, idx
);
3616 * Return whether there is a pagecache page to back given address within VMA.
3617 * Caller follow_hugetlb_page() holds page_table_lock so we cannot lock_page.
3619 static bool hugetlbfs_pagecache_present(struct hstate
*h
,
3620 struct vm_area_struct
*vma
, unsigned long address
)
3622 struct address_space
*mapping
;
3626 mapping
= vma
->vm_file
->f_mapping
;
3627 idx
= vma_hugecache_offset(h
, vma
, address
);
3629 page
= find_get_page(mapping
, idx
);
3632 return page
!= NULL
;
3635 int huge_add_to_page_cache(struct page
*page
, struct address_space
*mapping
,
3638 struct inode
*inode
= mapping
->host
;
3639 struct hstate
*h
= hstate_inode(inode
);
3640 int err
= add_to_page_cache(page
, mapping
, idx
, GFP_KERNEL
);
3644 ClearPagePrivate(page
);
3646 spin_lock(&inode
->i_lock
);
3647 inode
->i_blocks
+= blocks_per_huge_page(h
);
3648 spin_unlock(&inode
->i_lock
);
3652 static int hugetlb_no_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3653 struct address_space
*mapping
, pgoff_t idx
,
3654 unsigned long address
, pte_t
*ptep
, unsigned int flags
)
3656 struct hstate
*h
= hstate_vma(vma
);
3657 int ret
= VM_FAULT_SIGBUS
;
3665 * Currently, we are forced to kill the process in the event the
3666 * original mapper has unmapped pages from the child due to a failed
3667 * COW. Warn that such a situation has occurred as it may not be obvious
3669 if (is_vma_resv_set(vma
, HPAGE_RESV_UNMAPPED
)) {
3670 pr_warn_ratelimited("PID %d killed due to inadequate hugepage pool\n",
3676 * Use page lock to guard against racing truncation
3677 * before we get page_table_lock.
3680 page
= find_lock_page(mapping
, idx
);
3682 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3687 * Check for page in userfault range
3689 if (userfaultfd_missing(vma
)) {
3691 struct vm_fault vmf
= {
3696 * Hard to debug if it ends up being
3697 * used by a callee that assumes
3698 * something about the other
3699 * uninitialized fields... same as in
3705 * hugetlb_fault_mutex must be dropped before
3706 * handling userfault. Reacquire after handling
3707 * fault to make calling code simpler.
3709 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
,
3711 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3712 ret
= handle_userfault(&vmf
, VM_UFFD_MISSING
);
3713 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3717 page
= alloc_huge_page(vma
, address
, 0);
3719 ret
= PTR_ERR(page
);
3723 ret
= VM_FAULT_SIGBUS
;
3726 clear_huge_page(page
, address
, pages_per_huge_page(h
));
3727 __SetPageUptodate(page
);
3728 set_page_huge_active(page
);
3730 if (vma
->vm_flags
& VM_MAYSHARE
) {
3731 int err
= huge_add_to_page_cache(page
, mapping
, idx
);
3740 if (unlikely(anon_vma_prepare(vma
))) {
3742 goto backout_unlocked
;
3748 * If memory error occurs between mmap() and fault, some process
3749 * don't have hwpoisoned swap entry for errored virtual address.
3750 * So we need to block hugepage fault by PG_hwpoison bit check.
3752 if (unlikely(PageHWPoison(page
))) {
3753 ret
= VM_FAULT_HWPOISON
|
3754 VM_FAULT_SET_HINDEX(hstate_index(h
));
3755 goto backout_unlocked
;
3760 * If we are going to COW a private mapping later, we examine the
3761 * pending reservations for this page now. This will ensure that
3762 * any allocations necessary to record that reservation occur outside
3765 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3766 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3768 goto backout_unlocked
;
3770 /* Just decrements count, does not deallocate */
3771 vma_end_reservation(h
, vma
, address
);
3774 ptl
= huge_pte_lock(h
, mm
, ptep
);
3775 size
= i_size_read(mapping
->host
) >> huge_page_shift(h
);
3780 if (!huge_pte_none(huge_ptep_get(ptep
)))
3784 ClearPagePrivate(page
);
3785 hugepage_add_new_anon_rmap(page
, vma
, address
);
3787 page_dup_rmap(page
, true);
3788 new_pte
= make_huge_pte(vma
, page
, ((vma
->vm_flags
& VM_WRITE
)
3789 && (vma
->vm_flags
& VM_SHARED
)));
3790 set_huge_pte_at(mm
, address
, ptep
, new_pte
);
3792 hugetlb_count_add(pages_per_huge_page(h
), mm
);
3793 if ((flags
& FAULT_FLAG_WRITE
) && !(vma
->vm_flags
& VM_SHARED
)) {
3794 /* Optimization, do the COW without a second fault */
3795 ret
= hugetlb_cow(mm
, vma
, address
, ptep
, page
, ptl
);
3807 restore_reserve_on_error(h
, vma
, address
, page
);
3813 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3814 struct vm_area_struct
*vma
,
3815 struct address_space
*mapping
,
3816 pgoff_t idx
, unsigned long address
)
3818 unsigned long key
[2];
3821 if (vma
->vm_flags
& VM_SHARED
) {
3822 key
[0] = (unsigned long) mapping
;
3825 key
[0] = (unsigned long) mm
;
3826 key
[1] = address
>> huge_page_shift(h
);
3829 hash
= jhash2((u32
*)&key
, sizeof(key
)/sizeof(u32
), 0);
3831 return hash
& (num_fault_mutexes
- 1);
3835 * For uniprocesor systems we always use a single mutex, so just
3836 * return 0 and avoid the hashing overhead.
3838 u32
hugetlb_fault_mutex_hash(struct hstate
*h
, struct mm_struct
*mm
,
3839 struct vm_area_struct
*vma
,
3840 struct address_space
*mapping
,
3841 pgoff_t idx
, unsigned long address
)
3847 int hugetlb_fault(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
3848 unsigned long address
, unsigned int flags
)
3855 struct page
*page
= NULL
;
3856 struct page
*pagecache_page
= NULL
;
3857 struct hstate
*h
= hstate_vma(vma
);
3858 struct address_space
*mapping
;
3859 int need_wait_lock
= 0;
3861 address
&= huge_page_mask(h
);
3863 ptep
= huge_pte_offset(mm
, address
);
3865 entry
= huge_ptep_get(ptep
);
3866 if (unlikely(is_hugetlb_entry_migration(entry
))) {
3867 migration_entry_wait_huge(vma
, mm
, ptep
);
3869 } else if (unlikely(is_hugetlb_entry_hwpoisoned(entry
)))
3870 return VM_FAULT_HWPOISON_LARGE
|
3871 VM_FAULT_SET_HINDEX(hstate_index(h
));
3873 ptep
= huge_pte_alloc(mm
, address
, huge_page_size(h
));
3875 return VM_FAULT_OOM
;
3878 mapping
= vma
->vm_file
->f_mapping
;
3879 idx
= vma_hugecache_offset(h
, vma
, address
);
3882 * Serialize hugepage allocation and instantiation, so that we don't
3883 * get spurious allocation failures if two CPUs race to instantiate
3884 * the same page in the page cache.
3886 hash
= hugetlb_fault_mutex_hash(h
, mm
, vma
, mapping
, idx
, address
);
3887 mutex_lock(&hugetlb_fault_mutex_table
[hash
]);
3889 entry
= huge_ptep_get(ptep
);
3890 if (huge_pte_none(entry
)) {
3891 ret
= hugetlb_no_page(mm
, vma
, mapping
, idx
, address
, ptep
, flags
);
3898 * entry could be a migration/hwpoison entry at this point, so this
3899 * check prevents the kernel from going below assuming that we have
3900 * a active hugepage in pagecache. This goto expects the 2nd page fault,
3901 * and is_hugetlb_entry_(migration|hwpoisoned) check will properly
3904 if (!pte_present(entry
))
3908 * If we are going to COW the mapping later, we examine the pending
3909 * reservations for this page now. This will ensure that any
3910 * allocations necessary to record that reservation occur outside the
3911 * spinlock. For private mappings, we also lookup the pagecache
3912 * page now as it is used to determine if a reservation has been
3915 if ((flags
& FAULT_FLAG_WRITE
) && !huge_pte_write(entry
)) {
3916 if (vma_needs_reservation(h
, vma
, address
) < 0) {
3920 /* Just decrements count, does not deallocate */
3921 vma_end_reservation(h
, vma
, address
);
3923 if (!(vma
->vm_flags
& VM_MAYSHARE
))
3924 pagecache_page
= hugetlbfs_pagecache_page(h
,
3928 ptl
= huge_pte_lock(h
, mm
, ptep
);
3930 /* Check for a racing update before calling hugetlb_cow */
3931 if (unlikely(!pte_same(entry
, huge_ptep_get(ptep
))))
3935 * hugetlb_cow() requires page locks of pte_page(entry) and
3936 * pagecache_page, so here we need take the former one
3937 * when page != pagecache_page or !pagecache_page.
3939 page
= pte_page(entry
);
3940 if (page
!= pagecache_page
)
3941 if (!trylock_page(page
)) {
3948 if (flags
& FAULT_FLAG_WRITE
) {
3949 if (!huge_pte_write(entry
)) {
3950 ret
= hugetlb_cow(mm
, vma
, address
, ptep
,
3951 pagecache_page
, ptl
);
3954 entry
= huge_pte_mkdirty(entry
);
3956 entry
= pte_mkyoung(entry
);
3957 if (huge_ptep_set_access_flags(vma
, address
, ptep
, entry
,
3958 flags
& FAULT_FLAG_WRITE
))
3959 update_mmu_cache(vma
, address
, ptep
);
3961 if (page
!= pagecache_page
)
3967 if (pagecache_page
) {
3968 unlock_page(pagecache_page
);
3969 put_page(pagecache_page
);
3972 mutex_unlock(&hugetlb_fault_mutex_table
[hash
]);
3974 * Generally it's safe to hold refcount during waiting page lock. But
3975 * here we just wait to defer the next page fault to avoid busy loop and
3976 * the page is not used after unlocked before returning from the current
3977 * page fault. So we are safe from accessing freed page, even if we wait
3978 * here without taking refcount.
3981 wait_on_page_locked(page
);
3986 * Used by userfaultfd UFFDIO_COPY. Based on mcopy_atomic_pte with
3987 * modifications for huge pages.
3989 int hugetlb_mcopy_atomic_pte(struct mm_struct
*dst_mm
,
3991 struct vm_area_struct
*dst_vma
,
3992 unsigned long dst_addr
,
3993 unsigned long src_addr
,
3994 struct page
**pagep
)
3996 int vm_shared
= dst_vma
->vm_flags
& VM_SHARED
;
3997 struct hstate
*h
= hstate_vma(dst_vma
);
4005 page
= alloc_huge_page(dst_vma
, dst_addr
, 0);
4009 ret
= copy_huge_page_from_user(page
,
4010 (const void __user
*) src_addr
,
4011 pages_per_huge_page(h
), false);
4013 /* fallback to copy_from_user outside mmap_sem */
4014 if (unlikely(ret
)) {
4017 /* don't free the page */
4026 * The memory barrier inside __SetPageUptodate makes sure that
4027 * preceding stores to the page contents become visible before
4028 * the set_pte_at() write.
4030 __SetPageUptodate(page
);
4031 set_page_huge_active(page
);
4034 * If shared, add to page cache
4037 struct address_space
*mapping
= dst_vma
->vm_file
->f_mapping
;
4038 pgoff_t idx
= vma_hugecache_offset(h
, dst_vma
, dst_addr
);
4040 ret
= huge_add_to_page_cache(page
, mapping
, idx
);
4042 goto out_release_nounlock
;
4045 ptl
= huge_pte_lockptr(h
, dst_mm
, dst_pte
);
4049 if (!huge_pte_none(huge_ptep_get(dst_pte
)))
4050 goto out_release_unlock
;
4053 page_dup_rmap(page
, true);
4055 ClearPagePrivate(page
);
4056 hugepage_add_new_anon_rmap(page
, dst_vma
, dst_addr
);
4059 _dst_pte
= make_huge_pte(dst_vma
, page
, dst_vma
->vm_flags
& VM_WRITE
);
4060 if (dst_vma
->vm_flags
& VM_WRITE
)
4061 _dst_pte
= huge_pte_mkdirty(_dst_pte
);
4062 _dst_pte
= pte_mkyoung(_dst_pte
);
4064 set_huge_pte_at(dst_mm
, dst_addr
, dst_pte
, _dst_pte
);
4066 (void)huge_ptep_set_access_flags(dst_vma
, dst_addr
, dst_pte
, _dst_pte
,
4067 dst_vma
->vm_flags
& VM_WRITE
);
4068 hugetlb_count_add(pages_per_huge_page(h
), dst_mm
);
4070 /* No need to invalidate - it was non-present before */
4071 update_mmu_cache(dst_vma
, dst_addr
, dst_pte
);
4081 out_release_nounlock
:
4088 long follow_hugetlb_page(struct mm_struct
*mm
, struct vm_area_struct
*vma
,
4089 struct page
**pages
, struct vm_area_struct
**vmas
,
4090 unsigned long *position
, unsigned long *nr_pages
,
4091 long i
, unsigned int flags
, int *nonblocking
)
4093 unsigned long pfn_offset
;
4094 unsigned long vaddr
= *position
;
4095 unsigned long remainder
= *nr_pages
;
4096 struct hstate
*h
= hstate_vma(vma
);
4098 while (vaddr
< vma
->vm_end
&& remainder
) {
4100 spinlock_t
*ptl
= NULL
;
4105 * If we have a pending SIGKILL, don't keep faulting pages and
4106 * potentially allocating memory.
4108 if (unlikely(fatal_signal_pending(current
))) {
4114 * Some archs (sparc64, sh*) have multiple pte_ts to
4115 * each hugepage. We have to make sure we get the
4116 * first, for the page indexing below to work.
4118 * Note that page table lock is not held when pte is null.
4120 pte
= huge_pte_offset(mm
, vaddr
& huge_page_mask(h
));
4122 ptl
= huge_pte_lock(h
, mm
, pte
);
4123 absent
= !pte
|| huge_pte_none(huge_ptep_get(pte
));
4126 * When coredumping, it suits get_dump_page if we just return
4127 * an error where there's an empty slot with no huge pagecache
4128 * to back it. This way, we avoid allocating a hugepage, and
4129 * the sparse dumpfile avoids allocating disk blocks, but its
4130 * huge holes still show up with zeroes where they need to be.
4132 if (absent
&& (flags
& FOLL_DUMP
) &&
4133 !hugetlbfs_pagecache_present(h
, vma
, vaddr
)) {
4141 * We need call hugetlb_fault for both hugepages under migration
4142 * (in which case hugetlb_fault waits for the migration,) and
4143 * hwpoisoned hugepages (in which case we need to prevent the
4144 * caller from accessing to them.) In order to do this, we use
4145 * here is_swap_pte instead of is_hugetlb_entry_migration and
4146 * is_hugetlb_entry_hwpoisoned. This is because it simply covers
4147 * both cases, and because we can't follow correct pages
4148 * directly from any kind of swap entries.
4150 if (absent
|| is_swap_pte(huge_ptep_get(pte
)) ||
4151 ((flags
& FOLL_WRITE
) &&
4152 !huge_pte_write(huge_ptep_get(pte
)))) {
4154 unsigned int fault_flags
= 0;
4158 if (flags
& FOLL_WRITE
)
4159 fault_flags
|= FAULT_FLAG_WRITE
;
4161 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
;
4162 if (flags
& FOLL_NOWAIT
)
4163 fault_flags
|= FAULT_FLAG_ALLOW_RETRY
|
4164 FAULT_FLAG_RETRY_NOWAIT
;
4165 if (flags
& FOLL_TRIED
) {
4166 VM_WARN_ON_ONCE(fault_flags
&
4167 FAULT_FLAG_ALLOW_RETRY
);
4168 fault_flags
|= FAULT_FLAG_TRIED
;
4170 ret
= hugetlb_fault(mm
, vma
, vaddr
, fault_flags
);
4171 if (ret
& VM_FAULT_ERROR
) {
4175 if (ret
& VM_FAULT_RETRY
) {
4180 * VM_FAULT_RETRY must not return an
4181 * error, it will return zero
4184 * No need to update "position" as the
4185 * caller will not check it after
4186 * *nr_pages is set to 0.
4193 pfn_offset
= (vaddr
& ~huge_page_mask(h
)) >> PAGE_SHIFT
;
4194 page
= pte_page(huge_ptep_get(pte
));
4197 pages
[i
] = mem_map_offset(page
, pfn_offset
);
4208 if (vaddr
< vma
->vm_end
&& remainder
&&
4209 pfn_offset
< pages_per_huge_page(h
)) {
4211 * We use pfn_offset to avoid touching the pageframes
4212 * of this compound page.
4218 *nr_pages
= remainder
;
4220 * setting position is actually required only if remainder is
4221 * not zero but it's faster not to add a "if (remainder)"
4226 return i
? i
: -EFAULT
;
4229 #ifndef __HAVE_ARCH_FLUSH_HUGETLB_TLB_RANGE
4231 * ARCHes with special requirements for evicting HUGETLB backing TLB entries can
4234 #define flush_hugetlb_tlb_range(vma, addr, end) flush_tlb_range(vma, addr, end)
4237 unsigned long hugetlb_change_protection(struct vm_area_struct
*vma
,
4238 unsigned long address
, unsigned long end
, pgprot_t newprot
)
4240 struct mm_struct
*mm
= vma
->vm_mm
;
4241 unsigned long start
= address
;
4244 struct hstate
*h
= hstate_vma(vma
);
4245 unsigned long pages
= 0;
4247 BUG_ON(address
>= end
);
4248 flush_cache_range(vma
, address
, end
);
4250 mmu_notifier_invalidate_range_start(mm
, start
, end
);
4251 i_mmap_lock_write(vma
->vm_file
->f_mapping
);
4252 for (; address
< end
; address
+= huge_page_size(h
)) {
4254 ptep
= huge_pte_offset(mm
, address
);
4257 ptl
= huge_pte_lock(h
, mm
, ptep
);
4258 if (huge_pmd_unshare(mm
, &address
, ptep
)) {
4263 pte
= huge_ptep_get(ptep
);
4264 if (unlikely(is_hugetlb_entry_hwpoisoned(pte
))) {
4268 if (unlikely(is_hugetlb_entry_migration(pte
))) {
4269 swp_entry_t entry
= pte_to_swp_entry(pte
);
4271 if (is_write_migration_entry(entry
)) {
4274 make_migration_entry_read(&entry
);
4275 newpte
= swp_entry_to_pte(entry
);
4276 set_huge_pte_at(mm
, address
, ptep
, newpte
);
4282 if (!huge_pte_none(pte
)) {
4283 pte
= huge_ptep_get_and_clear(mm
, address
, ptep
);
4284 pte
= pte_mkhuge(huge_pte_modify(pte
, newprot
));
4285 pte
= arch_make_huge_pte(pte
, vma
, NULL
, 0);
4286 set_huge_pte_at(mm
, address
, ptep
, pte
);
4292 * Must flush TLB before releasing i_mmap_rwsem: x86's huge_pmd_unshare
4293 * may have cleared our pud entry and done put_page on the page table:
4294 * once we release i_mmap_rwsem, another task can do the final put_page
4295 * and that page table be reused and filled with junk.
4297 flush_hugetlb_tlb_range(vma
, start
, end
);
4298 mmu_notifier_invalidate_range(mm
, start
, end
);
4299 i_mmap_unlock_write(vma
->vm_file
->f_mapping
);
4300 mmu_notifier_invalidate_range_end(mm
, start
, end
);
4302 return pages
<< h
->order
;
4305 int hugetlb_reserve_pages(struct inode
*inode
,
4307 struct vm_area_struct
*vma
,
4308 vm_flags_t vm_flags
)
4311 struct hstate
*h
= hstate_inode(inode
);
4312 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4313 struct resv_map
*resv_map
;
4317 * Only apply hugepage reservation if asked. At fault time, an
4318 * attempt will be made for VM_NORESERVE to allocate a page
4319 * without using reserves
4321 if (vm_flags
& VM_NORESERVE
)
4325 * Shared mappings base their reservation on the number of pages that
4326 * are already allocated on behalf of the file. Private mappings need
4327 * to reserve the full area even if read-only as mprotect() may be
4328 * called to make the mapping read-write. Assume !vma is a shm mapping
4330 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4331 resv_map
= inode_resv_map(inode
);
4333 chg
= region_chg(resv_map
, from
, to
);
4336 resv_map
= resv_map_alloc();
4342 set_vma_resv_map(vma
, resv_map
);
4343 set_vma_resv_flags(vma
, HPAGE_RESV_OWNER
);
4352 * There must be enough pages in the subpool for the mapping. If
4353 * the subpool has a minimum size, there may be some global
4354 * reservations already in place (gbl_reserve).
4356 gbl_reserve
= hugepage_subpool_get_pages(spool
, chg
);
4357 if (gbl_reserve
< 0) {
4363 * Check enough hugepages are available for the reservation.
4364 * Hand the pages back to the subpool if there are not
4366 ret
= hugetlb_acct_memory(h
, gbl_reserve
);
4368 /* put back original number of pages, chg */
4369 (void)hugepage_subpool_put_pages(spool
, chg
);
4374 * Account for the reservations made. Shared mappings record regions
4375 * that have reservations as they are shared by multiple VMAs.
4376 * When the last VMA disappears, the region map says how much
4377 * the reservation was and the page cache tells how much of
4378 * the reservation was consumed. Private mappings are per-VMA and
4379 * only the consumed reservations are tracked. When the VMA
4380 * disappears, the original reservation is the VMA size and the
4381 * consumed reservations are stored in the map. Hence, nothing
4382 * else has to be done for private mappings here
4384 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
) {
4385 long add
= region_add(resv_map
, from
, to
);
4387 if (unlikely(chg
> add
)) {
4389 * pages in this range were added to the reserve
4390 * map between region_chg and region_add. This
4391 * indicates a race with alloc_huge_page. Adjust
4392 * the subpool and reserve counts modified above
4393 * based on the difference.
4397 rsv_adjust
= hugepage_subpool_put_pages(spool
,
4399 hugetlb_acct_memory(h
, -rsv_adjust
);
4404 if (!vma
|| vma
->vm_flags
& VM_MAYSHARE
)
4405 region_abort(resv_map
, from
, to
);
4406 if (vma
&& is_vma_resv_set(vma
, HPAGE_RESV_OWNER
))
4407 kref_put(&resv_map
->refs
, resv_map_release
);
4411 long hugetlb_unreserve_pages(struct inode
*inode
, long start
, long end
,
4414 struct hstate
*h
= hstate_inode(inode
);
4415 struct resv_map
*resv_map
= inode_resv_map(inode
);
4417 struct hugepage_subpool
*spool
= subpool_inode(inode
);
4421 chg
= region_del(resv_map
, start
, end
);
4423 * region_del() can fail in the rare case where a region
4424 * must be split and another region descriptor can not be
4425 * allocated. If end == LONG_MAX, it will not fail.
4431 spin_lock(&inode
->i_lock
);
4432 inode
->i_blocks
-= (blocks_per_huge_page(h
) * freed
);
4433 spin_unlock(&inode
->i_lock
);
4436 * If the subpool has a minimum size, the number of global
4437 * reservations to be released may be adjusted.
4439 gbl_reserve
= hugepage_subpool_put_pages(spool
, (chg
- freed
));
4440 hugetlb_acct_memory(h
, -gbl_reserve
);
4445 #ifdef CONFIG_ARCH_WANT_HUGE_PMD_SHARE
4446 static unsigned long page_table_shareable(struct vm_area_struct
*svma
,
4447 struct vm_area_struct
*vma
,
4448 unsigned long addr
, pgoff_t idx
)
4450 unsigned long saddr
= ((idx
- svma
->vm_pgoff
) << PAGE_SHIFT
) +
4452 unsigned long sbase
= saddr
& PUD_MASK
;
4453 unsigned long s_end
= sbase
+ PUD_SIZE
;
4455 /* Allow segments to share if only one is marked locked */
4456 unsigned long vm_flags
= vma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4457 unsigned long svm_flags
= svma
->vm_flags
& VM_LOCKED_CLEAR_MASK
;
4460 * match the virtual addresses, permission and the alignment of the
4463 if (pmd_index(addr
) != pmd_index(saddr
) ||
4464 vm_flags
!= svm_flags
||
4465 sbase
< svma
->vm_start
|| svma
->vm_end
< s_end
)
4471 static bool vma_shareable(struct vm_area_struct
*vma
, unsigned long addr
)
4473 unsigned long base
= addr
& PUD_MASK
;
4474 unsigned long end
= base
+ PUD_SIZE
;
4477 * check on proper vm_flags and page table alignment
4479 if (vma
->vm_flags
& VM_MAYSHARE
&&
4480 vma
->vm_start
<= base
&& end
<= vma
->vm_end
)
4486 * Search for a shareable pmd page for hugetlb. In any case calls pmd_alloc()
4487 * and returns the corresponding pte. While this is not necessary for the
4488 * !shared pmd case because we can allocate the pmd later as well, it makes the
4489 * code much cleaner. pmd allocation is essential for the shared case because
4490 * pud has to be populated inside the same i_mmap_rwsem section - otherwise
4491 * racing tasks could either miss the sharing (see huge_pte_offset) or select a
4492 * bad pmd for sharing.
4494 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4496 struct vm_area_struct
*vma
= find_vma(mm
, addr
);
4497 struct address_space
*mapping
= vma
->vm_file
->f_mapping
;
4498 pgoff_t idx
= ((addr
- vma
->vm_start
) >> PAGE_SHIFT
) +
4500 struct vm_area_struct
*svma
;
4501 unsigned long saddr
;
4506 if (!vma_shareable(vma
, addr
))
4507 return (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4509 i_mmap_lock_write(mapping
);
4510 vma_interval_tree_foreach(svma
, &mapping
->i_mmap
, idx
, idx
) {
4514 saddr
= page_table_shareable(svma
, vma
, addr
, idx
);
4516 spte
= huge_pte_offset(svma
->vm_mm
, saddr
);
4518 get_page(virt_to_page(spte
));
4527 ptl
= huge_pte_lock(hstate_vma(vma
), mm
, spte
);
4528 if (pud_none(*pud
)) {
4529 pud_populate(mm
, pud
,
4530 (pmd_t
*)((unsigned long)spte
& PAGE_MASK
));
4533 put_page(virt_to_page(spte
));
4537 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4538 i_mmap_unlock_write(mapping
);
4543 * unmap huge page backed by shared pte.
4545 * Hugetlb pte page is ref counted at the time of mapping. If pte is shared
4546 * indicated by page_count > 1, unmap is achieved by clearing pud and
4547 * decrementing the ref count. If count == 1, the pte page is not shared.
4549 * called with page table lock held.
4551 * returns: 1 successfully unmapped a shared pte page
4552 * 0 the underlying pte page is not shared, or it is the last user
4554 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4556 pgd_t
*pgd
= pgd_offset(mm
, *addr
);
4557 pud_t
*pud
= pud_offset(pgd
, *addr
);
4559 BUG_ON(page_count(virt_to_page(ptep
)) == 0);
4560 if (page_count(virt_to_page(ptep
)) == 1)
4564 put_page(virt_to_page(ptep
));
4566 *addr
= ALIGN(*addr
, HPAGE_SIZE
* PTRS_PER_PTE
) - HPAGE_SIZE
;
4569 #define want_pmd_share() (1)
4570 #else /* !CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4571 pte_t
*huge_pmd_share(struct mm_struct
*mm
, unsigned long addr
, pud_t
*pud
)
4576 int huge_pmd_unshare(struct mm_struct
*mm
, unsigned long *addr
, pte_t
*ptep
)
4580 #define want_pmd_share() (0)
4581 #endif /* CONFIG_ARCH_WANT_HUGE_PMD_SHARE */
4583 #ifdef CONFIG_ARCH_WANT_GENERAL_HUGETLB
4584 pte_t
*huge_pte_alloc(struct mm_struct
*mm
,
4585 unsigned long addr
, unsigned long sz
)
4591 pgd
= pgd_offset(mm
, addr
);
4592 pud
= pud_alloc(mm
, pgd
, addr
);
4594 if (sz
== PUD_SIZE
) {
4597 BUG_ON(sz
!= PMD_SIZE
);
4598 if (want_pmd_share() && pud_none(*pud
))
4599 pte
= huge_pmd_share(mm
, addr
, pud
);
4601 pte
= (pte_t
*)pmd_alloc(mm
, pud
, addr
);
4604 BUG_ON(pte
&& pte_present(*pte
) && !pte_huge(*pte
));
4609 pte_t
*huge_pte_offset(struct mm_struct
*mm
, unsigned long addr
)
4615 pgd
= pgd_offset(mm
, addr
);
4616 if (pgd_present(*pgd
)) {
4617 pud
= pud_offset(pgd
, addr
);
4618 if (pud_present(*pud
)) {
4620 return (pte_t
*)pud
;
4621 pmd
= pmd_offset(pud
, addr
);
4624 return (pte_t
*) pmd
;
4627 #endif /* CONFIG_ARCH_WANT_GENERAL_HUGETLB */
4630 * These functions are overwritable if your architecture needs its own
4633 struct page
* __weak
4634 follow_huge_addr(struct mm_struct
*mm
, unsigned long address
,
4637 return ERR_PTR(-EINVAL
);
4640 struct page
* __weak
4641 follow_huge_pmd(struct mm_struct
*mm
, unsigned long address
,
4642 pmd_t
*pmd
, int flags
)
4644 struct page
*page
= NULL
;
4647 ptl
= pmd_lockptr(mm
, pmd
);
4650 * make sure that the address range covered by this pmd is not
4651 * unmapped from other threads.
4653 if (!pmd_huge(*pmd
))
4655 if (pmd_present(*pmd
)) {
4656 page
= pmd_page(*pmd
) + ((address
& ~PMD_MASK
) >> PAGE_SHIFT
);
4657 if (flags
& FOLL_GET
)
4660 if (is_hugetlb_entry_migration(huge_ptep_get((pte_t
*)pmd
))) {
4662 __migration_entry_wait(mm
, (pte_t
*)pmd
, ptl
);
4666 * hwpoisoned entry is treated as no_page_table in
4667 * follow_page_mask().
4675 struct page
* __weak
4676 follow_huge_pud(struct mm_struct
*mm
, unsigned long address
,
4677 pud_t
*pud
, int flags
)
4679 if (flags
& FOLL_GET
)
4682 return pte_page(*(pte_t
*)pud
) + ((address
& ~PUD_MASK
) >> PAGE_SHIFT
);
4685 #ifdef CONFIG_MEMORY_FAILURE
4688 * This function is called from memory failure code.
4690 int dequeue_hwpoisoned_huge_page(struct page
*hpage
)
4692 struct hstate
*h
= page_hstate(hpage
);
4693 int nid
= page_to_nid(hpage
);
4696 spin_lock(&hugetlb_lock
);
4698 * Just checking !page_huge_active is not enough, because that could be
4699 * an isolated/hwpoisoned hugepage (which have >0 refcount).
4701 if (!page_huge_active(hpage
) && !page_count(hpage
)) {
4703 * Hwpoisoned hugepage isn't linked to activelist or freelist,
4704 * but dangling hpage->lru can trigger list-debug warnings
4705 * (this happens when we call unpoison_memory() on it),
4706 * so let it point to itself with list_del_init().
4708 list_del_init(&hpage
->lru
);
4709 set_page_refcounted(hpage
);
4710 h
->free_huge_pages
--;
4711 h
->free_huge_pages_node
[nid
]--;
4714 spin_unlock(&hugetlb_lock
);
4719 bool isolate_huge_page(struct page
*page
, struct list_head
*list
)
4723 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4724 spin_lock(&hugetlb_lock
);
4725 if (!page_huge_active(page
) || !get_page_unless_zero(page
)) {
4729 clear_page_huge_active(page
);
4730 list_move_tail(&page
->lru
, list
);
4732 spin_unlock(&hugetlb_lock
);
4736 void putback_active_hugepage(struct page
*page
)
4738 VM_BUG_ON_PAGE(!PageHead(page
), page
);
4739 spin_lock(&hugetlb_lock
);
4740 set_page_huge_active(page
);
4741 list_move_tail(&page
->lru
, &(page_hstate(page
))->hugepage_activelist
);
4742 spin_unlock(&hugetlb_lock
);